Executive Summary

This report surveys all known methods for obtaining transcripts of YouTube videos, focusing on their use to enrich the context available to large language models (LLMs). Transcripts—textual representations of a video’s spoken content—can greatly enhance information retrieval, question-answering, summarization, and other AI-driven tasks by converting audiovisual material into machine-readable text. We examine both native YouTube features (such as the built-in “Open transcript” UI and the official YouTube Data API) and external tools and techniques (including Python libraries, browser/work-around methods, and speech-recognition systems). We also discuss third-party services (human and AI transcription), real-world case studies, and the implications of using video transcripts in LLM pipelines. Throughout, we provide extensive detail, examples, and references:

• YouTube’s built-in transcript feature: Many videos on YouTube have manually uploaded captions or auto-generated transcripts accessible via the web interface (the “Show transcript” function) (Source: medium.com) (Source: www.notta.ai). However, this is not designed for programmatic use.
• YouTube Data API (Captions endpoint): The official YouTube Data API v3 provides a “captions” resource. Developers can list caption tracks for a video and download them in formats like SRT or “.vtt” (Source: developers.google.com) (Source: developers.google.com). This works for manual captions but not for auto-generated ones, unless they have been “published” by the video owner.
• TimedText (video.google.com): An undocumented HTTP endpoint accepts queries like http://video.google.com/timedtext?lang=en&v=VIDEOID to fetch transcripts without API keys (Source: stackoverflow.com). This only returns manually uploaded transcripts (auto-captions often require additional parameters) and yields XML-formatted results.
• Open-source libraries: Tools like youtube-transcript-api (Python) can scrape both manually provided and automatically generated captions without needing an API key (Source: github.com) (Source: github.com). Similarly, libraries such as pytube allow programmatic access to captions (e.g. YouTube(url).captions.get_by_language_code('en').generate_srt_captions()) (Source: stackoverflow.com). The yt-dlp command-line tool (with appropriate options or plugins) can also download transcripts and video metadata (Source: pypi.org).
• Speech-to-text (ASR) approaches: When no transcript exists, one can download the video’s audio (via tools like yt-dlp) and feed it into ASR systems. Modern ASR models range from open-source solutions (e.g. OpenAI’s Whisper (Source: forklog.com) to cloud platforms (Google Speech-to-Text, AWS Transcribe, etc.). OpenAI’s Whisper, trained on ~680,000 hours of multilingual audio, achieves near-human accuracy and supports many languages (Source: medium.com) (Source: forklog.com). Commercial APIs (Google, Microsoft, Rev.ai, DeepGram, etc.) also support transcript generation in dozens of languages (Source: www.summaraize.com).
• Other techniques: Even manual or semi-automated workarounds exist. One can use Google Docs’ Voice Typing by playing the video audio into it (Source: www.summaraize.com), or online tools (such as YouTubeTranscript.com, Notta, or SummarAIze) that either fetch embedded captions or perform ASR on the fly (Source: www.notta.ai) (Source: www.summaraize.com).
• Applications with LLMs: Transcripts enable LLM-based processing of video content. For example, pipelines using LangChain or LlamaIndex can fetch a video’s transcript, chunk it, and feed it to an LLM for summarization or QA (Source: www.toolify.ai) (Source: abvijaykumar.medium.com). Case studies illustrate how transcripts are used for tasks like automatic chaptering (Source: medium.com) (Source: medium.com) and conversational QA on lecture videos (Source: abvijaykumar.medium.com). Research even shows transcripts being used as data – e.g. an analysis of 740k hours of YouTube talks found language shifts attributable to ChatGPT usage (Source: huggingface.co) (Source: huggingface.co).
• Quality and Practical Considerations: Most methods yield timestamped segments lacking punctuation and requiring cleanup (Source: medium.com) (Source: stackoverflow.com). Automatically generated captions often contain errors and misspeaking; manually created transcripts are more accurate but less common. Legislative and copyright concerns also arise since transcripts are derivative works whose copyright is held by the video owner (Source: insight7.io).
• Future Directions: As video content grows, improving transcript extraction is crucial. LLMs with larger context (and emerging multimodal models) may directly ingest transcripts. Better ASR models and legal frameworks will shape how transcripts enrich AI systems.

Overall, this report provides an exhaustive overview of all recognized methods to obtain YouTube video transcripts, together with analysis of their merits, use cases, and future prospects. Citations from official documentation, academic work, and industry sources substantiate each claim.

Introduction

YouTube is a vast repository of audiovisual knowledge, hosting billions of videos across every domain. However, by default YouTube (and other video platforms) make the spoken content of videos inaccessible to text-based systems like large language models (LLMs) except through their captions or transcripts. Converting video to text (speech-to-text) is therefore a critical step for tasks such as question answering, summarization, content analysis, and data retrieval using LLMs or indexing systems. Transcripts also support accessibility (e.g. for deaf users) and search engine indexing of video content (Source: gotranscript.com) (Source: www.captioncut.com).YouTube itself offers automatic captioning for many videos and allows content creators to upload closed captions (manually crafted transcripts). These transcripts can sometimes be accessed by viewers via the “Open transcript” menu in the YouTube player UI. However, our goal is programmatic access: “All the different ways to get the transcript of YouTube videos” implies methods suitable for automation and integration with LLM pipelines, not just manual copying.

This report thoroughly surveys these methods, ranging from official APIs and endpoints provided by Google/YouTube, to third-party tools and services, to speech-recognition approaches that bypass YouTube’s own transcripts entirely. We review the technical procedures, the quality and format of resulting transcripts, and case studies illustrating how transcripts empower AI workflows. We consider both broad categories (like “use the YouTube Data API”) and specific tools (like the youtube-transcript-api Python package) where relevant. We also discuss the contextual purposes of transcripts: how they are ingested into LLM contexts (e.g., with Retrieval-Augmented Generation) and what implications this has.

The remainder of this report is organized as follows. First, we detail native YouTube features for transcripts (the UI and official API). Next, we examine community-developed libraries and browser tricks for scraping transcripts. Then, we cover speech-to-text methods (including offline and cloud ASR solutions). We follow with sections on use in LLMs (including data pipelines and case studies) and implications/future trends. Each section includes in-depth analysis, examples, data and citations. Tables summarize key tools/methods for quick reference. All facts are backed by sources, including YouTube’s own documentation, developer blog posts, research findings, and industry reports.

Methods for Obtaining YouTube Video Transcripts

Various approaches exist to obtain the transcript (the text of spoken content) from a YouTube video. Broadly, these can be grouped into (1) native YouTube mechanisms, (2) specialized software libraries/tools, and (3) speech-to-text transcription. We examine each category in detail, highlighting specific techniques within them.

1. Native YouTube Mechanisms

1.1 YouTube’s “Show transcript” UI (Desktop/Mobile)

Description: Many YouTube videos have captions (closed captions or subtitles) that can be opened by the user in the web player. On desktop, this is accessed via the three-dot menu → “Show transcript”. The transcript panel then appears, usually on the right, showing time-coded text (Source: www.notta.ai). This includes auto-generated captions (if the video owner has enabled them) or user-uploaded captions. On mobile, the “Show transcript” option also exists under the video menu in many cases (Source: www.notta.ai).

Usage: This is a manual process: a user must physically open the transcript panel and copy text. It can be useful for ad-hoc viewing or copying small segments. For example, Notta’s guide explains how to scroll to “Show transcript” under the video description and then copy the text into a document (Source: www.notta.ai). One must disable timestamps if not needed (the UI often shows them by default).

Pros:

• No technical setup required. It works out-of-the-box on any video that has captions.
• Immediate demonstration. Good for quickly inspecting a transcript.

Cons:

• Not scalable or automated. It is manual; not suitable for feeding transcripts into software.
• Limited to what’s available. If the video has no captions (auto or manual), this menu won’t show.
• Quality issues. The displayed transcript often lacks punctuation and may display partial sentences or filler (“ums”). Captions may be poorly aligned to sentences (Source: medium.com).
• UI constraints. YouTube’s interface may truncate very long lines or omit certain elements. Copy-paste may include timestamps or require toggling to remove them.

Because of these drawbacks, most programmatic solutions bypass the UI and access transcripts via other interfaces.

1.2 YouTube Data API – Captions Resource

Description: YouTube provides an official Data API (v3) for developers to programmatically interact with YouTube data. Within this API, the Captions resource allows listing, uploading, updating, and downloading caption tracks associated with a video (Source: developers.google.com). Each “caption” resource corresponds to one language track (manual subtitle file) on a specific video.

How it works: To use this API, one must obtain OAuth or API credentials and have permission (the video’s owner usually) to access the captions. The key steps are:

• List caption tracks: Call captions.list with a videoId. The response lists the available caption tracks for that video (usually just the manual ones; it does not return the actual text (Source: developers.google.com). Each track includes metadata (language, kind, etc.).
• Download captions: Given a caption track ID from above, call captions.download. This returns the captions file, typically in its original format (e.g. “.srt” or “.vtt”), unless requested otherwise (Source: developers.google.com). You can specify tfmt (text format) or tlang (target language) parameters to change it.

For example, Google’s documentation shows that captions.download can retrieve a caption track in a specified format and language (Source: developers.google.com).

Sources: The official API docs clearly outline the captions resource and its methods (Source: developers.google.com) (Source: developers.google.com). For example, Google’s documentation notes: “The captions resource includes a snippet with details like the videoId, language, trackKind, … The caption track’s snippet.isAutoSynced indicates if the track is time-synced” (Source: developers.google.com). It also explicitly calls out the captions.download method (“the caption track is returned to its original format” unless parameters specify otherwise (Source: developers.google.com).

Pros:

• Official support: As part of YouTube’s API, it is documented and stable (subject to updates by Google).
• Structured results: You get well-formatted outputs (SRT, VTT, or text).
• Capabilities: You can get multiple languages if they exist, and even translate captions via the API.
• Legal compliance: Using the official API respects YouTube’s terms.

Cons:

• Permissions/Quota: Requires an API key or OAuth credentials with youtube.force-ssl scopes (Source: developers.google.com). Also subject to YouTube’s quota limits, which might constrain bulk downloads.
• No auto captions: It apparently only accesses captions that have been uploaded or provided by the user, not the auto-generated tracks (Source: stackoverflow.com). This is a major limitation: many videos only have auto captions available (and the API does not list those as caption tracks). For instance, a 2014 StackOverflow thread notes “none of the solutions…retrieve automatically generated subtitles… I came up with github.com/jdepoix/youtube-transcript-api” (Source: stackoverflow.com), implying the Data API cannot directly fetch auto-captions.
• Tied to video owner: You can only download tracks for a video if you have access (admin, same account, etc.). You cannot arbitrarily fetch captions from any video via the API unless it’s public captions (which may still require special calls).
• Complex setup: For simple use-cases, setting up OAuth and making HTTP requests is more involved than some open-source tools.

1.3 Google’s TimedText Endpoint

Description: Aside from the official API, there exists an undocumented HTTP endpoint that can return YouTube transcripts via a simple URL query. This endpoint is video.google.com/timedtext, which predates YouTube API v3. It accepts query parameters for video ID and language, such as:

http://video.google.com/timedtext?lang=en&v=<VIDEO_ID>

This returns captions (in XML format) if a transcript is available in that language.

How it works: As noted by community sources, one can issue a GET request to the above URL with the YouTube video ID and language code to fetch the transcript text. For example, a top answer on StackOverflow instructs: “Just do a GET request on: http://video.google.com/timedtext?lang={LANG}&v={VIDEOID}. You don't need any api/oauth/etc. to access this.” (Source: stackoverflow.com).

Behavior: Typically, this returns the manually-provided caption track. For auto-generated captions (“asr”), a separate parameter &track=asr may be needed (though in practice this often fails). A comment in the same StackOverflow thread indicates that auto-generated captions require track=asr and still did not work in one case (Source: stackoverflow.com). The library youtube-transcript-api (below) was in part created because this timedtext method did not handle auto-captions by itself (Source: stackoverflow.com).

Pros:

• No API key needed: It is a straightforward HTTP GET.
• Simplicity: Good for quick scripts or embedding in other tools.

Cons:

• Manual captions only: By default it returns only non-auto captions. As per StackOverflow reports, using track=asr to get auto captions often fails (Source: stackoverflow.com).
• Raw output: The XML is relatively simple (each <text start="..." dur="...">…</text>) but still requires parsing. It may not include nice formatting.
• Undocumented: Because it’s not an official API, Google could change or shut it down at any time without notice.
• Limited to one language per request: You have to know the language code, or loop through possibilities to find available languages.

1.4 YouTube Live Captions

A related note: YouTube Live streams also have automatic live captions. These can sometimes be accessed via similar APIs (e.g. if live captioning is turned on, the captions resource may list them). Additionally, there exist realtime caption WebSocket streams (undocumented). However, since the question focuses on “transcripts of YouTube videos” in general, live streams are beyond its main scope.

2. Community Tools and Libraries

Given the limitations of YouTube’s own interfaces, many developers and companies have created tools to retrieve transcripts. These often combine web scraping, public endpoints, and ASR to function without needing official API credentials.

2.1 youtube-transcript-api (Python)

One of the most widely used libraries is youtube-transcript-api (by jdepoix). It is a Python package available on PyPI (Source: github.com). Key features:

• No API key needed: It scrapes transcripts using public endpoints.
• Supports auto-captions: Crucially, it can fetch transcripts even if they were auto-generated by YouTube.
• Multiple languages: It can list available transcripts and fetch in specific languages, as well as translate them.
• Output format: It returns a list of dictionaries, each with text, start, and duration keys for each caption snippet.
• Community maintained: Over 650 forks on GitHub, MIT-licensed.

Example usage is simple:

from youtube_transcript_api import YouTubeTranscriptApi
transcript = YouTubeTranscriptApi.get_transcript("ErnWZxJovaM", languages=["en"])

This returns e.g.:

[
 {'text': '[Music]', 'start': 1.17, 'duration': 9.11},
 {'text': 'good afternoon everyone and welcome to', 'start': 10.28, 'duration': 2.60},
 {'text': 'MIT 6.S191 my name is Alexander Amini', 'start': 12.88, 'duration': 3.96},
 ...
]

(Excerpt adapted from Le Borgne, 2024 (Source: medium.com).)

The GitHub README highlights: “This is a python API which allows you to retrieve the transcript/subtitles for a given YouTube video. It also works for automatically generated subtitles…” (Source: github.com) (Source: github.com). Crucially, the project explicitly notes it “does not require a headless browser” or API key (Source: github.com), setting it apart from Selenium-based scrapers.

Pros:

• Ease of use: Simple Python calls.
• Handles auto-captions: A big advantage over the official Data API method.
• Language handling: Can download or translate transcripts.
• Open-source: MIT License, active GitHub repo.

Cons:

• No punctuation: The returned text has no punctuation, all lowercase (typical of YouTube auto captions) (Source: medium.com). Post-processing is needed for readability.
• Dependent on YouTube’s site code: If YouTube changes how transcripts are served, the library may break (though it is actively maintained).
• Python only: Directly useful in Python applications (though one could call it via subprocess).

Le Borgne (2024) provides an example of using this library to fetch transcripts for an MIT lecture video (Source: medium.com). He notes that the raw output “lacks punctuation and contains typos” (Source: medium.com). For instance, he observes transcripts like 'MIT sus1 191' instead of ‘MIT 6.S191’. This exemplifies the typical imperfections of raw caption text.

2.2 pytube (Python)

Pytube is a popular Python library for downloading YouTube videos and metadata. It also provides access to caption tracks.

• Example flow (from StackOverflow) (Source: stackoverflow.com):

  from pytube import YouTube
  yt = YouTube("https://www.youtube.com/watch?v=wjTn_EkgQRg")
  caption = yt.captions.get_by_language_code('en')
  srt_text = caption.generate_srt_captions()
  print(srt_text)
  

  This code fetches the English captions and formats them in SRT style.

The StackOverflow snippet shows using get_by_language_code('en') and then generate_srt_captions() (Source: stackoverflow.com). The library can also list available captions via yt.captions.keys(). Note that old versions of pytube may have bugs, but current versions generally work.

Pros:

• No API key: Similar to youtube-transcript-api, it scrapes.
• Outputs SRT/XML: generate_srt_captions() yields text with numbering and timecodes.
• Part of a larger toolkit: If you are already using Pytube to download video or audio, you can get captions in the same library.

Cons:

• Only manual captions: Pytube’s captions getter typically only sees user-uploaded caption tracks, not auto-generated ones. (I.e., it likely wraps the official API behind scenes; it won’t fetch “asr” tracks by default.)
• No punctuation fix: SRT will still have no added punctuation beyond what’s in the captions.
• Python dependency: Again, requires Python environment.

2.3 yt-dlp and youtube-dl (CLI/Python)

youtube-dl and its active fork yt-dlp are command-line tools (with Python libraries) to download YouTube content. They support downloading videos, audio, metadata, and captions.

One can retrieve transcripts with yt-dlp via:

• --write-auto-sub or --write-sub: Options that download English (or specified language) subtitles, in formats like .srv1 or .vtt. For example: yt-dlp --write-auto-sub --sub-lang en --get-sub <video URL>.
• Python scripts: There are wrappers and plugins (like the yt-dlp-transcripts PyPI package) that automate batch retrieval of transcripts for videos, channels, or playlists (Source: pypi.org).

The PyPI package yt-dlp-transcripts advertises itself as “a Python tool for extracting video information and transcripts… built on yt-dlp and youtube-transcript-api” (Source: pypi.org). It supports single videos, entire playlists, and channels, and can export transcripts to CSV (Source: pypi.org). This indicates that behind the scenes it integrates both yt-dlp (for basic extraction) and youtube-transcript-api (for transcripts).

Pros:

• Bulk processing: Can handle playlists and multiple videos with progress tracking (Source: pypi.org).
• Metadata: Not only transcripts, but also titles, descriptions, views, and more can be extracted in one go.
• Flexible: CLI and Python APIs available.

Cons:

• Setup required: yt-dlp needs to be installed, and depending on the method, might need FFmpeg or other codecs if doing audio extraction.
• Maintenance issues: YouTube often changes its internal APIs, which occasionally breaks youtube-dl/yt-dlp until patched.
• Quality of captions: Still depends on what captions exist (for --write-auto-sub, it grabs auto generaed subs by environment).
• No punctuation fix: As always, output raw segments.

2.4 Web and Browser Extensions

Several browser extensions and web tools enable direct retrieval of YouTube transcripts:

• Chrome/Firefox Extensions: For example, Tactiq (an “AI meeting tool”) has a “YouTube Summaries” or caption grabbing feature. These often work by injecting scripts to parse the YouTube UI. (Tactiq’s own blog FAQ suggests using Python etc., but the Chrome plugin does it directly (Source: tactiq.io).) Since such tools often use the same underlying endpoints as youtube-transcript-api, they share similar pros/cons (they require user activation, can fetch transcripts programmatically).

• Online Services: Websites like YouTubeTranscript.com, DownSub.com, or SubtitleCat.com allow you to paste a YouTube URL and often provide the transcript as plain text. These typically just wrap the timedtext endpoint or call youtube-transcript-api in the backend. For instance, SummarAIze’s blog notes “Websites like YouTubeTranscript.com provide free transcription services. You enter the video URL, and they generate a transcript” (Source: www.summaraize.com). DeepGram’s free demo can generate transcripts for videos (Source: www.summaraize.com).

• Google Docs Voice Typing: A clever trick is to open Google Docs in Chrome, activate “Voice typing” under Tools, and play the YouTube video audio into your microphone (possibly at high volume or using stereo mix). Google Docs will attempt to transcribe in real-time (Source: www.summaraize.com). This requires a quiet environment and yields only as good an OCR transcript as the speech recognition, but can be done for free without coding.

• Screen Recording to Text: In absence of any tools, one could simply record the screen/stream and then run that audio through any transcription tool. This is essentially the ASR approach discussed in section 3.

Pros:

• No coding required: Many such tools are user-friendly.
• ASR-backed options: Some (like Notta (Source: www.notta.ai) or SummarAIze) claim to use advanced ASR to improve on YouTube’s auto captions.

Cons:

• Inconsistency: Quality and features vary widely. Free sites may not always work reliably, or may require signup.
• Terms of use: Some may not respect YouTube’s terms or copyright restrictions.
• Privacy: Pasting a URL sends data to a third party.
• Costs: Premium features may require payment (e.g. Notta’s advanced editing).

Overall, these browser/web methods are most useful for quick single-videos or non-technical users, rather than large-scale data pipelines.

3. Speech-to-Text (ASR) Approaches

When no satisfactory transcript is available from YouTube itself, one can generate a transcript by running the video’s audio through automatic speech recognition (ASR). This can be done using:

• Download video/audio then transcribe: First, download the video or its audio track (e.g. using yt-dlp or the YouTube API), then feed the audio into an ASR engine.
• Cloud ASR APIs: Services like Google Cloud Speech-to-Text, AWS Transcribe, Azure Speech, IBM Watson, DeepGram, Rev AI, etc. accept audio input (or video URL/stream) and return captions.
• Open-source ASR: Engines like OpenAI Whisper (and forks like faster-whisper), Mozilla DeepSpeech, Coqui STT, Kaldi, etc. The OpenAI Whisper model in particular has become very popular because it is open-source, highly accurate, and supports many languages (Source: forklog.com) (Source: medium.com).

3.1 Workflow for ASR Transcription

A typical pipeline (for Python, say) is:

1. Obtain audio from video. For example, using yt-dlp:

   yt-dlp -x --audio-format wav https://www.youtube.com/watch?v=VIDEOID
   

   or via Python: yt_dlp.YoutubeDL(...).extract_info(video_url, download=True) with appropriate options. This yields an audio file (e.g. VIDEOID.wav).

2. Transcription. Pass the audio file to the ASR model or API. For instance, with OpenAI’s Whisper (using faster-whisper for speed) (Source: medium.com):

   from faster_whisper import WhisperModel
   model = WhisperModel("large-v3", device="cuda", compute_type="float16")
   segments, info = model.transcribe("VIDEOID.wav", initial_prompt="Add punctuation.", language="en")
   

   This outputs segments containing text, start, and end timestamps (Source: medium.com) (Source: medium.com).

3. Post-processing. Many ASR outputs lack punctuation or have mistakes. One can optionally run a text post-processor (sometimes using an LLM) to format and correct the transcript (Source: medium.com). Le Borgne (2024) notes that Whisper’s output added punctuation (significantly improving readability) compared to the raw YouTube auto-transcript (Source: medium.com), though slight errors remained (e.g. “MIT Success 191” instead of “MIT 6.S191”).

4. Integration. Now the transcript (a plain text string or list of segments) can be fed into an LLM pipeline. It may need splitting into chunks (due to token limits) (Source: www.toolify.ai) (Source: medium.com).

3.2 Example: OpenAI Whisper

OpenAI released Whisper in 2022 as a state-of-the-art open-source ASR (Source: forklog.com). According to OpenAI, Whisper was trained on 680,000 hours of multilingual data, enabling it to handle accents, noise, and jargon (Source: forklog.com). It supports dozens of languages. Critical properties (from the GitHub README and announcements):

• Multilingual: e.g. English, Spanish, Chinese, etc.
• High accuracy: Near “human-level” robustness on many tasks, particularly in Whisper’s larger model variants (Source: forklog.com).
• Open-source (MIT): Can be run locally (no API costs).
• Model sizes: Ranging from small (faster, less accurate) to large (“large-v3” being most accurate, 50 GB download). Faster-whisper or other forks optimize speed on GPUs (Source: medium.com).
• Uses: Researchers and engineers frequently apply Whisper to transcribe YouTube videos. For example, the blog by Devang Tomar (2023) demonstrates using Whisper to transcribe a TED-Ed video: first extracting audio with yt-dlp, then running Whisper and (optionally) sending the transcript to GPT-3 for summarization (Source: medium.com).

Whisper’s performance is a step up from baseline YouTube captions. Le Borgne (2024) compares Whisper’s “large-v3” output against YouTube’s auto-captions for a lecture. Whisper added punctuation and generally improved readability. But some errors (like mis-recognizing a course code) still occurred (Source: medium.com). Nonetheless, Whisper’s results, combined with its free availability, make it a powerful tool for transcript generation.

3.3 Commercial ASR APIs

Cloud providers offer speech-to-text services that can directly accept audio or video URLs:

• Google Cloud Speech-to-Text: Recognizes 125 languages/dialects. Known for integration with Google’s ecosystem.
• AWS Transcribe: Amazon’s ASR, with features like speaker diarization.
• Microsoft Azure Speech: Another enterprise option with 85+ languages.
• Rev AI: The AI arm of the Rev transcription service, supports many languages and possibly custom dictionary.
• DeepGram: Offers an API for real-time and batch transcription (advertised free tier supports up to 30 languages (Source: www.summaraize.com).
• YouTube’s own ASR: Note that using YouTube’s auto-captions itself is just leveraging Google’s ASR, but they do not expose it beyond what we discussed.

These APIs typically charge by minute of audio. They often yield nice transcripts with punctuation (though sometimes errors). Many are used in media indexing, research, and accessibility. For instance, Summaraize mentions DeepGram: “Free and fast way to generate a transcript from a YouTube video in over 30 languages” (Source: www.summaraize.com).

Pros of ASR approach:

• Language coverage: Can handle videos with no captions or in languages where YouTube’s auto-caption is poor or absent.
• Quality: State-of-the-art models can exceed YouTube’s auto-subtitle quality, especially with noise or multiple speakers.
• Control: You can choose model (fast vs accurate), specify accent hints, translator, etc.
• Scalability: Can automate retrieval for any video.

Cons:

• Compute/Cost: Running Whisper large locally or paying a cloud per minute can be significant for huge video collections.
• Time: Transcribing hours of video takes time (Whisper large takes ~4x real-time on a good GPU (Source: medium.com).
• No enhancement of content: Like YouTube’s transcript, ASR transcript is “just text” – any meaning beyond words is not captured.
• License/copyright: If using someone else’s video to generate a transcript, legal issues apply (see later).

In sum, ASR is a catch-all method: it will work for any video (assuming clear audio), whereas other methods rely on transcripts being provided. Often, a hybrid approach is used: first attempt to fetch an existing transcript (to save work/cost), and fallback to ASR if none is found.

3.4 ASR Performance and Accuracy

Substantial research exists on ASR accuracy. Generally, the Word Error Rate (WER) of state-of-the-art models can be on the order of a few percent on clean speech, but increases with noise, accents, or poor audio. User reports suggest YouTube’s auto-captions (as of 2023) can vary widely in accuracy (some news reports claim up to ~90% error in the worst cases, though rigorous stats are scarce). By contrast, Whisper’s largest models often achieve single-digit WER on benchmark tasks, even with background noise (Source: forklog.com).

For example, a citizen study by Cisdem (June 2025) found varying accuracies by language and speaker clarity, but found Whisper much better than baseline auto subtitles. (They report that Whisper’s WER is near 5–10% on well-recorded English speeches, while YouTube’s auto-caption had WER above 15–20% for many utterances (Source: www.transcribetube.com).) (Note: this is a blog, not formal study, but illustrates the trend that dedicated ASR is superior to rudimentary auto-captions.)

Modern ASR also supports multiple speakers or diarization, punctuation, and sometimes recognition of extended vocabulary. In practice, human transcripts are still more accurate, but ASR provides a cost-effective alternative, especially when millions of videos are involved.

4. Quality, Formats, and Limitations of Transcripts

Regardless of method, the raw transcripts often share common limitations:

• Lack of Punctuation/Grammar: YouTube auto-captions and many ASR outputs omit punctuation, produce run-on text, and have spelling/grammar errors (Source: medium.com) (Source: stackoverflow.com). For example, Le Borgne found YouTube’s transcript for an academic talk had no punctuation and mis-transcribed “6.S191” as “sus1 191” (Source: medium.com).
• Timestamps and Segmentation: Most transcripts (from all sources) are chunked into short phrases with timestamps. This is useful for referencing timing, but undesired if one just needs plain text. For LLM feeding, one typically strips timestamps or merges segments into paragraphs.
• Error Rate: Automatic transcripts contain mis-recognitions, especially with technical terms, names, accents, overlapping speakers, or low audio quality. Even Whisper has occasional mistakes (e.g., “MIT Success 191” instead of “MIT 6.S191” (Source: medium.com).
• Language Support: Some videos have multiple subtitle tracks (e.g. English auto-captions plus a Spanish translation). Not all tools retrieve all languages by default. “youtube-transcript-api” can list multiple available languages, for instance.
• Length and Context Window: Long videos produce very long transcripts. LLM context windows (even the longest models) have limits (e.g., 32k or 100k tokens). This requires smart chunking and retrieval strategies (Source: www.toolify.ai) (Source: medium.com).
• Copyright/Permission: Transcripts are usually considered derivative works of the video. The video owner typically holds rights to both the audio and any manually created captions (Source: insight7.io). Using public captions may be allowed, but automated extraction tools must still adhere to the Terms of Service. We discuss legal implications next.

Despite these drawbacks, transcripts remain invaluable data. The act of transforming spoken words into text “enriches” the content for LLMs, allowing advanced NLP techniques to be applied.

5. Case Studies and Applications

Beyond generic methods, it helps to see how transcripts are used in practice. Here are a few representative case studies and examples drawn from literature and practice:

• Academic Lecture Indexing: Yann-Aël Le Borgne (2024) processed the transcript of an MIT deep learning lecture (with an MIT license) using LLMs and TF-IDF to automatically generate video chapter headings (Source: medium.com) (Source: medium.com). His workflow began by retrieving the YouTube transcript (using youtube-transcript-api) (Source: medium.com), then post-processing it into paragraphs, and finally splitting into chapters. This kind of semantically structured output is only possible because the audio was turned into text.

• Subtitle Generation and Enhancement: Summarization tools like SummarAIze (2024) highlight using YouTube transcripts as a basis for content repurposing (Source: www.summaraize.com). Companies offering video AI (e.g. Verbit, Rev, CaptionCut) leverage transcripts to improve SEO, accessibility, and user engagement. As one marketing article notes, captioned videos increased viewer completion by 80% (Source: www.captioncut.com), indicating a strong push for transcript accuracy and completeness.

• Conversational Q&A (RAG): Vijay Kumar (2024) demonstrates a RAG chatbot using LlamaIndex: he uses the YoutubeTranscriptReader (built on youtube-transcript-api) to fetch a video’s transcript and index it. Then the LLM can answer questions about the video’s content (Source: abvijaykumar.medium.com). He emphasizes that the implementation is “very simple”: “using the youtube_transcript_api to extract the transcript… and use that to create the index” (Source: abvijaykumar.medium.com). This exemplifies how transcripts become the knowledge base for LLMs.

• Video Summarization with LangChain: A tutorial explains using LangChain’s youtube_loader to fetch transcripts, then running an OpenAI LLM (e.g. GPT-3 or GPT-4) to summarize (Source: www.toolify.ai). An important note is splitting long transcripts for token limits (Source: www.toolify.ai). It shows that transcripts can feed directly into load_summarize_chain to produce concise summaries (Source: www.toolify.ai).

• Cultural Linguistics Study: A large-scale research project analyzed 740,249 hours of YouTube academic talks transcripts to study the influence of ChatGPT on human speech (Source: huggingface.co). Surprisingly, they detected statistically significant shifts in vocabulary (“delve”, “comprehend”, “boast”, etc.) after ChatGPT’s release (Source: huggingface.co) (Source: huggingface.co). This case shows transcripts being treated as data for sociolinguistic analysis, made possible only because tens of thousands of videos were transcribed (via some large-scale method, presumably an ASR pipeline or by using owner-provided captions).

• Educational Use: Researchers have noted the value of transcripts for E-learning. For example, Lichera (2019) discusses how transcripts help second-language learners, linguistic analysis, and video search (Source: gotranscript.com). (Our report’s scope is technical, but pedagogically, transcripts aid comprehension and note-taking.)

• Accessibility Compliance: Many platforms now require transcripts for accessibility (e.g., US CVAA mandates captions on online videos). Thus, transcripts can often be found through institutional channels. While not a “method” per se, this legal landscape increases the availability of transcripts in education and public sectors.

These examples illustrate the diverse uses of YouTube transcripts once obtained: from summarization and Q&A to corpus linguistics. They motivate why so many methods exist to get transcripts in the first place.

6. Legal and Ethical Considerations

Transcripts, being text derived from audio/video, implicate copyright law and platform policies. Key points:

• Copyright: According to authoritative sources, a transcription of copyrighted video is itself a derivative work encompassed by the original’s copyright (Source: insight7.io). YouTube further states that uploaded captions belong to the video owner. Thus, downloading and using transcripts (even auto-generated ones) potentially requires permission, especially for redistribution or commercial use. Working with transcripts “for personal study” or fair use may be permissible, but broad use can risk infringement. From Insight7 (2023): “YouTube’s automatic captioning…transcripts for videos…are considered derivative works… the copyright of the transcription belongs to the video owner, not YouTube” (Source: insight7.io).

• YouTube Terms of Service: Programmatic retrieval of transcripts must comply with YouTube’s TOS. The official API method obviously does. Scraping via unpaid endpoints (video.google.com/timedtext) is unofficial and may contravene site scraping rules. Using downloaded audio with Whisper is more clearr: the transcripts are user-generated content, so one must respect the original content’s license. Many free YouTube videos are provided under licensces (e.g. CC-BY-NC) that allow internal use.

• Privacy: If videos contain personal information or private conversations, transcribing them raises privacy concerns. This is more an issue if one shares transcripts of private videos, but even public livestreams could have individuals unexpectedly captured.

• Bias and Errors: Auto-transcripts can misgender or misrepresent speakers (e.g., mis-labeling names or accents). LLMs downstream might hallucinate or emphasize mis-transcribed content. Ethically, one must be cautious that biases in ASR (e.g., lower accuracy for some dialects) don’t propagate into model outputs.

In practice, source code authors and tools often add disclaimers. For example, the Insight7 article warns creators to review tools’ terms and ensure compliance (Source: insight7.io). Similarly, any production LLM service using transcripts should document data lineage and obtain proper rights.

Integration with LLMs: Use of Transcripts to Enrich Context

After obtaining a transcript text, the next step is integrating it into the LLM’s pipeline. This section discusses how transcripts are leveraged to “enrich LLM context”, following modern patterns like Retrieval-Augmented Generation (RAG), fine-tuning, prompt engineering, etc.

7.1 Retrieval-Augmented Generation (RAG) with Transcripts

RAG architectures enhance LLM answers with external knowledge. For YouTube content, transcripts are a natural “knowledge store”. A typical flow is:

1. Indexing transcripts: The transcript (plain text) is segmented (e.g. into paragraphs or chunks of ~1000 words). Each chunk is embedded (via a vector model) and stored in a vector database.
2. User query: A user asks a question related to the video’s content.
3. Retrieval: The system finds the most semantically similar transcript chunks to the query.
4. ARM with LLM: The retrieved chunks are concatenated and fed to the LLM as context (often with a system prompt), and the LLM generates an answer.

This paradigm is exemplified by LangChain and LlamaIndex tools. For instance, LangChain’s YouTubeLoader (from recently added functionality) can load a YouTube URL’s transcript and convert it into documents automatically. The Toolify blog shows code using youtube_loader.from_youtube_url(...) followed by loader.load() to get a list of documents, each containing text and metadata (Source: www.toolify.ai). Those docs can be summarized or passed into chains.

Vijay Kumar’s LlamaIndex example (2024) details using YoutubeTranscriptReader to extract the transcript then building an index. In his words: “We will be using the youtube_transcript_api to extract the transcript from a YouTube video, use that to create the index”(for RAG) (Source: abvijaykumar.medium.com). This shows transcripts feeding straight into the RAG indexing pipeline.

Benefits: Using transcripts fills knowledge gaps for the LLM. The model then answers from that specific content (instead of hallucinating). This is especially useful for factual questions about a video (“What experiment did the lecturer demonstrate?”, “What conclusion did the CEO mention?”, etc.). It transforms the LLM into a Q&A system on video data.

Challenges: Transcript lengths often exceed token limits, so chunking and retrieval (as above) is essential. Also, transcripts may contain noise (fillers, irrelevant tangents), so embeddings and retrieval must be tuned to handle that. Moreover, if the video content covers multiple topics, a simple keyword search on the transcript can guide to the relevant part.

7.2 Summarization and Question-Answering

Even without an interactive query, transcripts can feed summarization pipelines. As an example, LangChain’s load_summarize_chain can take the entire transcript (or pieces) and return a textual summary. The Toolify article illustrates using diagram = load_summarize_chain(model) then result = summary_chain.run(transcript) to get a concise summary (Source: www.toolify.ai).

Similarly, one can fine-tune or prompt an LLM to produce structured notes or key points from a transcript. Some third-party apps (like YouTube summarizers) do this to generate video notes.

This use of transcripts is a form of context injection: it enriches the prompt with relevant information extracted from the video, rather than relying on the LLM’s pretrained knowledge (which might not include the video specifics). Chatbots like ChatGPT often struggle with "private knowledge" from a video unless given its transcript.

LangChain also notes a practical limit: if the transcript is very long, exceeding the model’s context window, one must split it. For example, in one pipeline, the transcript was split via a “recursive character splitter” to fit within token constraints (Source: www.toolify.ai). Another guide notes that GPT-4o-mini handles ~5000 characters well, while Llama-3 8B can handle only ~1500, necessitating careful chunking (Source: medium.com).

7.3 Hybrid Approaches

In some cases, transcripts are used in combination with other modalities:

• Video+Transcript Q&A: Vision-language LLMs (like GPT-4 Vision) can process short video clips or key frames, but for long videos transcripts are still needed. Some new research attempts to directly answer questions from video without transcripts (by analyzing audio/speech with LLMs), but this is nascent. For now, transcripts remain the primary bridge to the audio content.

• Caption translation: If a video’s transcript is in one language, it can be machine-translated (via models or APIs) into another, then fed to LLM. Tools like youtube-transcript-api even support on-the-fly translation of transcripts (via Google Translate) (Source: stackoverflow.com).

• Integration with analytics: Some companies link transcripts with video analytics (sentiment, speaker ID, topics) to drive content recommendation. This is beyond LLMs, but is another “enrichment” use case.

7.4 Real-world Example: YouTube Chatbot

To illustrate an end-to-end case: Suppose we want a chatbot that answers questions about a popular science YouTube lecture. We might do:

• Use youtube-transcript-api to pull the English transcript (since the creator enabled auto-captions). This yields 3,000 words in timestamp blocks.
• Clean and combine into paragraphs.
• Split into 8 chunks of ~400 tokens each, then embed each chunk into a Pinecone/Weaviate vector database.
• User asks: “What is the main conclusion of the lecture?” The system embeds this query and retrieves the top 2 chunks most relevant.
• The LLM (e.g. GPT-4o) is prompted with: “According to the following transcript excerpts from the lecture by [speakers], answer the question…” followed by the retrieved text. The model outputs a precise answer.
• Behind the scenes, we cite relevant excerpts with 【時間 stamp】 if needed for sources.

This workflow is a practical manifestation of the RAG pattern and yields an “LLM with video knowledge”. The key component was obtaining the transcript.

Implications and Future Directions

Increased Accessibility and Record-Keeping

The abundance of transcripts (from automated methods) will further democratize access to video content. Researchers can perform text-based search across videos; accessibility tools can provide captions in multiple languages. In the future, platforms may integrate live AI summarization or highlight generation from transcripts to aid navigation.

Multi-modal LLMs

LLMs are rapidly evolving to absorb multi-modal inputs (images, audio). Some vision-language models aim to process video directly. However, the relative ease of dealing with text means transcripts will remain crucial for some time. It is possible that future LLMs may internally transcribe video themselves (blurring the line), but currently the clarified transcript helps too.

Legal and Ethical Frameworks

As transcripts are used more for model training and deployment, clearer guidelines will emerge. For example, auto-generated captions might become part of a video’s metadata and be licensed similarly. Researchers and companies might need standardized disclaimers when using scraped transcripts.

Improved Tools and Accuracy

We expect continued improvements in ASR (e.g. Whisper-like models getting better, specialized models for noisy content, etc.). Specialized transcript tools may add features like speaker diarization (identifying “Speaker 1/2”), sentiment tags, or hyperlinks to video timeline. LLMs themselves might be fine-tuned to polish transcripts, adding punctuation or clarifying ambiguous terms, as was hinted by the “initial_prompt” trick with Whisper (Source: medium.com).

Large-Scale Video Corpora

Datasets of YouTube transcripts (like YT-20M) are being built for research (Source: huggingface.co). This may enable training LLMs on spoken-expressed content. The cross-pollination of human language and AI language in these transcripts, as seen by the ChatGPT vocabulary shift (Source: huggingface.co), may accelerate ongoing cultural changes.

Model and Context Windows

One constraint is context window size. As pointed out, transcripts for a 1-hour lecture (10k+ words) exceed even the largest model contexts. Future LLM architectures may allow millions of tokens, reducing the need for chunking. Alternately, hierarchical models could first compress transcripts (TL;DR style) before ingestion.

Integrating Real-Time Transcripts

Live streams on YouTube already have real-time auto-captions. Soon, one can imagine on-the-fly LLM analysis of live transcripts (e.g. a bot summarizing a live event every minute). The tools to do that (streaming ASR + LLM) are on the horizon.

Data Table Summaries

To aid comparison, we present two summary tables:

Table 1: Methods for Obtaining YouTube Transcripts (advantages/disadvantages).

Table 2: Example ASR Models/Services (approximate capabilities).

(ASR comparison data compiled from vendor documentation and industry sources.)

Data Analysis and Observations

While this report is qualitative, some quantitative context underscores the importance of transcripts:

• Video Consumption: YouTube users watch billions of hours monthly. According to Statista, YouTube users watched over 1 billion hours of video per day in 2018 (Source: gotranscript.com) (likely higher now). Captions greatly enhance this content’s utility.
• Caption Usage: Surveys indicate captions are widely used. For instance, 80% of viewers are more likely to complete a video if captions are available (Source: gotranscript.com), and captioned videos get 40% more views on average (Source: www.captioncut.com). This suggests demand for transcripts beyond mere compliance.
• Language Reach: In accessibility and SEO, converting speech to text indexes every word. One SEO report notes that search bots “can’t ‘hear’ videos” but can index transcript text (Source: gotranscript.com). Given YouTube’s role as a major search platform, transcripts multiply the “searchability” of content by orders of magnitude.
• LLM Context: Modern LLMs like GPT-4 have context windows up to ~32k tokens (or more in new models) (Source: www.toolify.ai). A one-hour video (~10k words) thus fits within a single pass of GPT-4o (1M context). This opens the practical possibility of entirely ingesting a video transcript into one model prompt (with minimal chunking). The fact that frameworks mention “token limits” implies many transcripts do exceed those windows and must be chunked【49†L103-L109†61†L12-L17】. Efficient retrieval pipelines therefore often use transcript segments as independent documents.

Implications and Future Directions

Obtaining YouTube transcripts is not just a technical exercise: it has broader implications:

• AI Advancement: Transcripts feed world knowledge into AI. As a user, if one asks GPT-4 about content from a recent video, the model’s answer quality now hinges on whether that video’s text can be provided. Methods to fetch transcripts thus have real impact on AI-driven information access.

• Long Documents in LLMs: As context windows expand, it becomes feasible to input longer transcripts directly. Models might one day process entire documentaries in one go. This suggests future LLMs might have built-in pipelines to ingest transcripts.

• Multimodal Trends: In the future, we may see integrated pipelines: e.g., directly extracting transcripts (via joint audio-text models) and summarizing them on-the-fly as video plays. YouTube or social platforms might offer built-in AI summaries using their own ASR+LLM.

• Standardization of Transcripts: There may be standardized metadata on how transcripts are distributed (e.g. embedding transcript URLs or files in video metadata). This would make retrieval easier and legal.

• Privacy and Security: As more transcripts become available, privacy for speakers is a concern. AI-generated transcripts might inadvertently capture personal data from videos. Systems will need privacy filtering (e.g., automatically anonymizing personal identifiers in transcripts).

• Benchmarking and Evaluation: The AI community could develop benchmarks for video transcription quality or pipelines (like creating multimodal QA datasets from video+transcript). Indeed some research (e.g. TVQA tasks) already combines video and transcripts for evaluation.

• Educational Uses: Particularly for educational content (lectures, tutorials), transcripts enable note-taking apps, flashcard generation, or comprehension analytics. The synergy of transcripts and LLMs could transform online learning.

• Multilingual and Cross-lingual: With advances in translation and multilingual ASR, one could retrieve a transcript in one language and translate it to another on the fly, instantly making foreign-language content accessible to a global LLM.

Overall, transcripts bridge the gap between visual media and text-based AI. Efforts to refine transcript extraction (for accuracy, cost, and coverage) will continue to be crucial as we push LLMs to encompass more real-world data. Our survey has shown that many tools are already available, and even more may emerge, to ensure that “whatever can be said on YouTube, can be read and understood by an LLM.”

Conclusion

In this report we have comprehensively catalogued every known approach to obtaining the transcript of a YouTube video for use in large language models. We covered:

• Native YouTube Features: the “Show transcript” UI and the official Data API’s captions resource (Source: developers.google.com). These methods rely on the video having captions in the first place.
• Public Endpoints and Scraping: the undocumented timedtext endpoint (Source: stackoverflow.com), and open-source libraries (e.g. youtube-transcript-api (Source: github.com), pytube (Source: stackoverflow.com) that scrape transcripts, often retrieving even auto-generated subtitles.
• Third-Party Tools: browser extensions, web apps, and services like Notta or DeepGram (which boast high accuracy across many languages (Source: www.notta.ai) (Source: www.summaraize.com).
• Automatic Speech Recognition: downloading the audio and using ASR systems (notably OpenAI Whisper (Source: forklog.com) among others) to produce transcripts with high fidelity.
• Integration Strategies: pipelines to feed transcripts into LLMs (via RAG/Q&A (Source: abvijaykumar.medium.com) (Source: www.toolify.ai), summarization tools (Source: www.toolify.ai), and analysis tasks (Source: huggingface.co) (Source: huggingface.co).
• Case Studies: practical examples from chapter generation (Source: medium.com) to Q&A chatbots (Source: abvijaykumar.medium.com) demonstrate the utility of transcripts in AI workflows.
• Challenges: issues of accuracy, formatting (lack of punctuation, timestamps (Source: medium.com) (Source: stackoverflow.com), language coverage, model context limits (Source: www.toolify.ai), and legal constraints (Source: insight7.io) were discussed.

In each section, we provided evidence-based analysis with dozens of citations. For instance, the YouTube API documentation (Source: developers.google.com) (Source: developers.google.com), GitHub libraries (Source: github.com), developer blogs (Source: medium.com) (Source: www.toolify.ai), and research findings (Source: huggingface.co) (Source: huggingface.co) all underpin our discussion. Tables summarize capabilities and trade-offs at a glance.

Key takeaways: There is no single “best way” – the choice depends on factors like video origin, desired accuracy, development resources, and licensing. It is often prudent to first try an official or open approach (YouTube API, timedtext, youtube-transcript-api) to save cost, and then fall back to transcribing audio with ASR if needed. The ecosystem offers options for both casual use and industrial pipelines.

Future outlook: As video continues to dominate online information, methods to convert it to text will grow in importance. We anticipate refinements in ASR, more integrated programming interfaces, and innovative AI tools (like summarizers and question-answering systems) built directly around transcripts. The synergy between video content and LLMs will only deepen.

In summary, any robust AI project seeking to “read” YouTube videos should consider all the methods detailed here. By leveraging transcripts – through YouTube’s own features, clever programming, or ASR – one can significantly enrich an LLM’s context and enable powerful new capabilities.

References

• Google YouTube Data API – Captions Resource (Methods: list, download) (Source: developers.google.com) (Source: developers.google.com).
• StackOverflow – retrieving transcripts via YouTube APIs/endpoints (Source: stackoverflow.com) (Source: stackoverflow.com).
• GitHub – youtube-transcript-api (Python) (Source: github.com) (Source: github.com).
• Yann-Aël Le Borgne (2024), “Automate Video Chaptering with LLMs and TF-IDF” (Medium) (Source: medium.com) (Source: medium.com).
• Wikipedia – YouTube (platform) entry (cited via dev site or stats).
• StackOverflow – using pytube to download captions (Source: stackoverflow.com).
• PyPI – yt-dlp-transcripts project (Source: pypi.org).
• Notta Blog (2024), “How to Get a YouTube Transcript…” (Source: www.notta.ai).
• SummarAIze Blog (2023), “How to get the transcript of a YouTube video…” (Source: www.summaraize.com).
• Insight7 (2023), “YouTube Transcription and Copyright” (Source: insight7.io).
• Toolify (2024), “Unlocking the Power of YouTube Transcripts with LangChain” (Source: www.toolify.ai) (Source: www.toolify.ai).
• Hugging Face – “Empirical evidence of LLM’s influence on human spoken language” (Source: huggingface.co) (Source: huggingface.co).
• OpenAI (2022) – “Whisper: Robust Speech Recognition”, press (via ForkLog) (Source: forklog.com).
• Cisdem (2025) – Blog, “YouTube Auto Caption Accuracy Test” (citing Verizon Media stats) (Source: gotranscript.com).
• CaptionCut (2025) – “Why Video Captions Are Essential… 2025” (industry stats) (Source: www.captioncut.com).
• Le Borgne (2024), code examples and evaluation of Whisper vs YouTube transcripts (Source: medium.com).
• Vijay Kumar (2024, Medium) – “Retrieval Augmented Generation (RAG) — Chatbot for YouTube with LlamaIndex” (Source: abvijaykumar.medium.com).
• Pereira et al. (2023) – Hugging Face Daily Papers summary of video-language models (YT-20M dataset) (Source: huggingface.co).
• Various documentation (YouTube API, Whisper GitHub) and tool READMEs.

(All sources above are cited inline; bracketed numbers refer to those IV. tool references.)