Salin

Inspect, resolve, download, and clean YouTube subtitle tracks from Python.

Salin is a typed Python library for working with YouTube subtitle tracks. It reports which manual and automatic tracks a video offers, resolves requested languages against configurable match and source policies, downloads captions in seven formats, and can return transcripts in memory as snippets. Auto-generated captions can also be cleaned into deduplicated prose.

• Manual and automatic tracks, including regional variants such as en-US and pt-BR
• vtt, srt, json3, srv1, srv2, srv3, and ttml
• Files on disk with download(), snippets in memory with fetch()
• Cleaned transcripts with clean_transcript()
• Bulk downloads through a thread pool
• Rate limits reported as RateLimitedError and DownloadStatus.RATE_LIMITED
• Browser cookies, cookies.txt, and HTTP/SOCKS proxies
• Fully typed (py.typed), Python 3.12+

Install

pip install salin

Or with uv:

uv add salin

Requires Python 3.12 or newer.

Quick start

from salin import Salin

client = Salin(languages="en")

result = client.download("https://www.youtube.com/watch?v=dQw4w9WgXcQ")
for sub in result.ok_subs:
    print(sub.sub_path)          # salin_downloads/dQw4w9WgXcQ.en.vtt

transcript = client.fetch("dQw4w9WgXcQ")
for snippet in transcript:
    print(snippet.start, snippet.text)

Subtitles are written to ./salin_downloads unless you pass output_dir to the client or to the call.

Inspecting available tracks

inspect() reports the manual and automatic tracks a video exposes, along with the formats each track is available in. Nothing is downloaded.

url = "https://www.youtube.com/watch?v=dQw4w9WgXcQ"
availability = Salin().inspect(url)

availability.video_id
availability.title
availability.manual          # frozenset({'de-DE', 'en', 'es-419', 'ja', 'pt-BR'})
availability.auto            # frozenset({'en', 'en-orig'})
availability.manual_formats  # {'en': frozenset({'json3', 'srt', 'srv1', ...}), ...}
availability.auto_formats

Resolving languages

resolve() maps each requested language to a track according to language_match, source_policy, and optional per-language aliases.

resolved = Salin().resolve(
    url,
    languages={"en": ["en-US", "en-GB"], "de": ["de-DE", "de-AT"]},
    language_match="exact",
    source_policy="manual_only",
)

for item in resolved:
    print(item.requested, item.resolved, item.source)

The defaults are language_match="regional" and source_policy="manual_then_auto". Available values:

LanguageMatch.EXACT        # "en" resolves only to "en"
LanguageMatch.REGIONAL     # "en" may resolve to en-US, en-orig, etc.

SubtitleSourcePolicy.MANUAL_ONLY
SubtitleSourcePolicy.AUTO_ONLY
SubtitleSourcePolicy.MANUAL_THEN_AUTO
SubtitleSourcePolicy.AUTO_THEN_MANUAL

The two settings constrain each other. For example, language_match="exact" with source_policy="auto_only" will not fall back to a manual en track even when one exists. A request that cannot be satisfied resolves to None.

Downloading subtitles

client = Salin(languages=["en", "de"], output_dir="subs")
result = client.download(url)

for sub in result.subtitles:
    print(sub.requested, sub.resolved, sub.source, sub.status, sub.sub_path)

Each requested language becomes its own file, named <video_id>.<language>.<format>. VTT is the default; pick another format per call:

client.download(url, subtitle_format="srt")

Per-subtitle failures are part of the result rather than exceptions. status is ok, missing, unavailable, rate_limited, or error, with details in .error; result.ok_subs returns only the files that were written. Pass strict=True to raise instead.

Fetching transcripts in memory

fetch() returns snippets directly and writes nothing to disk. It accepts a full URL or a bare video ID and honors the same language, source, and proxy options as download().

transcript = Salin(languages="en").fetch("dQw4w9WgXcQ")

len(transcript)                 # number of snippets
transcript.video_id             # 'dQw4w9WgXcQ'
transcript.language             # 'English' (YouTube's name for the track)
transcript.language_code        # 'en'
transcript.is_generated         # False for manual, True for auto
transcript.text                 # every snippet joined with a space
transcript.to_raw_data()        # [{'text': ..., 'start': ..., 'duration': ...}]

for snippet in transcript:      # .text / .start / .duration in seconds
    print(snippet.start, snippet.text)

fetch() reads the json3 track when one exists, falling back to srv3, srv1, srv2, ttml, or vtt.

Migrating from youtube-transcript-api

- from youtube_transcript_api import YouTubeTranscriptApi
- transcript = YouTubeTranscriptApi().fetch(video_id, languages=("en",))
+ from salin import Salin
+ transcript = Salin(languages="en").fetch(video_id)

The returned objects mirror the ones from youtube-transcript-api, with a few differences:

• The language list belongs on the client or in a languages= override, not in a positional argument.
• Snippets are normalized: inline markup, speaker chevrons (>>), and stage directions such as [Music] are stripped, and whitespace is collapsed.
• preserve_formatting=True is not supported and raises SalinValueError.
• Failures raise Salin errors: TranscriptUnavailableError when no requested language resolves, and SubtitleDownloadError when the caption request itself fails. A 429 raises RateLimitedError with .retry_after set when the server sent a Retry-After header.

Cleaning auto-captions

Downloaded files are exactly what YouTube serves: auto-generated tracks come in short overlapping windows, with word-timing tags and display markup in between. clean_transcript() merges the overlaps and strips the markup.

multiple hours to capture the same level
multiple hours to capture the same level
of<00:48:19.119><c> mistake</c><00:48:19.520><c> because</c> ...
of mistake because it requires like

After cleaning:

multiple hours to capture the same level of mistake because it requires like

from salin import Salin, clean_transcript

result = Salin(languages="en", subtitle_format="vtt").download(url)
paths = clean_transcript(result, output_dir="transcripts")

To clean a file without writing anything:

from salin import clean_cues

subtitle = result.ok_subs[0]
text = clean_cues(subtitle.sub_path, subtitle.format)

max_workers cleans in parallel (default 1). Without output_dir, each transcript is written next to its source. Every subtitle format is parsed: vtt, srt, json3, srv1, srv2, srv3, and ttml.

VTT is approximate. It is the only rolling format, and YouTube sometimes repeats a word inside a single cue where the overlap merge cannot see it. On a 72-minute auto-captioned podcast the cleaned VTT was 0.3% longer (41 words) than the same video's other six formats. Manual tracks are usually not rolling; prefer srt, srv1, srv3, json3, or ttml when you need exact text.

Command line

Installing the package also installs a salin command; python -m salin works if the script directory is not on PATH.

salin inspect dQw4w9WgXcQ
salin fetch dQw4w9WgXcQ                        # caption text on stdout
salin fetch dQw4w9WgXcQ --format srt --out talk.srt
salin download -i ids.txt -l en -l de -w 3 -o subs
salin clean subs/ --recursive --output-dir transcripts/
salin doctor --check dQw4w9WgXcQ

fetch and download accept positional URLs or --input FILE, default to en, and share the language, cookie, and proxy options. inspect --json, download --json, and fetch --format jsonl produce machine-readable output; exit codes are 0 for success, 1 for partial or failed runs, 2 for usage errors, 3 when rate-limited, and 4 when no transcript exists.

Every flag, JSON schema, exit-code edge case, and batch recipe is documented in CLI.md.

Bulk downloads

download_many() is an Iterator[SubtitleResult] backed by a thread pool. Results are yielded as they finish, so order is preserved only with max_workers=1.

client = Salin(languages="en")

for result in client.download_many(urls, max_workers=3):
    if not result.processable:
        continue
    for subtitle in result.ok_subs:
        print(subtitle.sub_path)

result.processable is a copy of the result containing only the successful subtitles, or None when nothing succeeded.

How many workers to use

More workers shorten a batch until YouTube's per-request latency starts rising with concurrency. Measured on 72 videos that already had English captions, split into six blocks of 12. Every worker count ran every block twice in randomized order, and each speedup is paired against the one-worker run of the same block (12 paired measurements per count):

We tested 1, 2, 3, and 6 workers; 4 and 5 were not measured. Three is the clear knee at roughly 2.3x sequential. Six was the fastest count tested, but its advantage over three was not reliable across the 12 paired blocks: a single slow request can add tens of seconds to any block, and that variance dominates the remaining gap. Use three workers as a default and six when a batch needs to finish sooner.

The number to watch is throughput while the batch is running. Even two workers sustain about 22 videos/min and six about 37/min, both above the 11–15 videos/min that completed long runs without a block. Workers make a batch finish sooner; they do not make a high request rate safe. For unattended runs, cap the overall rate by pausing the consumer between chunks:

import time

for i, result in enumerate(client.download_many(urls, max_workers=6), 1):
    # ...
    if i % 25 == 0:
        time.sleep(60)

The generator stops submitting new work while the consumer sleeps, so in-flight requests drain during the pause. At six workers, 25 videos take roughly 40s, making each cycle about 100s, or ~15 videos/min.

These timings count only the block runs; gaps and cooldowns were excluded.

Rate limits

Salin adds no delay by default (sleep_interval_subtitles=0). Raise it when request volume, not latency, is the concern:

client = Salin(languages=["en", "de", "ja"], sleep_interval_subtitles=2)

Each extra language then costs (languages - 1) x interval seconds per video. sleep_interval_requests is forwarded to yt-dlp and only applies when yt-dlp performs the download itself.

YouTube rate-limits its caption endpoint independently of page extraction. When that happens Salin records a distinct status instead of retrying blindly:

result = client.download(url)

if result.is_rate_limited:
    print(result.retry_after)          # largest Retry-After seen, or None
    for sub in result.rate_limited_subs:
        print(sub.requested, sub.error)

• A 429 is never retried and never delegated to yt-dlp, which would retry the same limited endpoint with long, silent backoff.
• Retry-After (delta-seconds or an HTTP-date) is parsed into SubtitleFile.retry_after; SubtitleResult.retry_after reports the largest value across the requested subtitles.
• Transient failures (5xx, connection errors, timeouts) are retried twice with jittered backoff. An expired signed URL (403) is refreshed once by re-extracting the video.

Use strict=True to raise instead of returning a partial result:

from salin import RateLimitedError

try:
    client.download(url, strict=True)
except RateLimitedError as exc:
    wait = exc.retry_after or 60

For batches, max_consecutive_rate_limits stops the generator once that many rate-limited results arrive in a row. The counter resets on any other result; with max_workers > 1, "consecutive" is approximate because results arrive in completion order.

for result in client.download_many(urls, max_consecutive_rate_limits=3):
    ...

Cookies and proxies

Some videos only expose subtitles to signed-in, region-locked, or differently routed clients. Configure this on the client or per call:

client = Salin(
    languages="en",
    cookies="firefox:myprofile",        # read from a browser
    # cookiefile="~/cookies.txt",       # or a Netscape cookies.txt you control
    # proxy="socks5://127.0.0.1:1080",  # or route through a proxy
)

result = client.download(url, cookies="chrome", proxy="http://127.0.0.1:8080")

cookies uses yt-dlp's syntax: BROWSER[+KEYRING][:PROFILE][::CONTAINER], for example "chrome", "firefox", "firefox+gnomekeyring", or "chromium::Container 1". cookiefile must be an existing Netscape-format cookies.txt. proxy accepts http, https, socks4, socks4a, socks5, and socks5h URLs. Invalid values raise SalinValueError before any request is made.

These settings apply to metadata extraction, the direct caption fetch, the yt-dlp fallback, fetch(), and the re-extraction after an expired signed URL. inspect() and resolve() accept them too.

fetch() reads caption bytes with urllib, so it only supports http:// proxies. With https://, socks4, or socks5, it raises SalinValueError and points at download(), which delegates the fetch to yt-dlp and supports every scheme.

Security note: cookies reads the browser's cookie store and sends your logged-in session to the remote site with every request. Only use it with sources you trust, and prefer a cookies.txt you control on shared machines.

Performance

Metadata extraction dominates: roughly 4s per video, against ~0.3s for the caption bytes. inspect(), resolve(), fetch(), and download() each extract once, so calling three of them on the same video repeats that cost three times. download() already resolves every requested language and reports the results, so most flows need a single call.

To share one extraction across stages, pass info=:

from salin import extract_info

info = extract_info(url)
availability = client.inspect(url, info=info)
resolved = client.resolve(url, languages=["en", "de"], info=info)
transcript = client.fetch(url, languages="en", info=info)
result = client.download(url, languages=["en", "de"], info=info)

That pipeline takes about 4.6s instead of 12s. info must come from extract_info(); hand-built dicts are rejected and playlists are not supported.

extract_info() keeps one yt-dlp client per thread and request configuration, so back-to-back calls skip about 0.3s of setup; a repeated extraction drops from ~3.0s to ~1.2–1.6s. Salin also caps yt-dlp's default retry counts and socket timeout, so a stuck request fails fast rather than stalling a batch.

How it compares

If you only need in-memory snippets from YouTube, youtube-transcript-api is smaller and faster per video although you get blocked pretty much almost immediately (under a hundred videos) if you don't have residential proxies (or any of the solutions they document in their docs). Salin covers more of the surrounding work: files on disk, format conversion, cleaning, bulk runs, and behavior you can reason about when YouTube starts refusing requests.

Benchmarks

Measured on one residential IP (Windows) with yt-dlp 2026.07.04, youtube-transcript-api 1.2.4, and Salin 0.1.0, English only, sequential, no artificial delay. The machine was in normal use, so sub-second differences are noise; block onset and success/failure are reliable.

Random sample: 1,000 videos

The sample is 1,000 videos drawn uniformly from a database of 9,776. Only 50 of them were marked as having captions, so the sample is not biased toward videos known to have subtitle tracks. Every video was attempted:

• Zero blocks over 88.1 minutes, an average of 11.4 videos/min.
• Latency on the 898 successes: mean 5.58s, p50 5.06s, p95 9.66s.
• Throughput rose through the run: 9.8 → 12.9 videos/min per 200-video segment.
• The other 101 videos were re-extracted and classified. None had a native English track that Salin failed to fetch: 96 have no captions at all, and 5 are non-English videos whose only en track is a machine translation, which Salin skips by default rather than silently returning translated text.

An earlier 480-video run over videos already marked as captioned averaged 15.0 videos/min with zero blocks (p50 3.25s). The higher rate reflects a different, easier sample rather than a code change.

captions_available is not a reliable filter: 898 of these 1,000 videos had captions, but only 50 were marked. Treat an unset flag as unknown.

Against youtube-transcript-api

On the 55 videos both libraries handled, content matched exactly: median raw characters 7,032 vs 7,032, median snippets 116 vs 116, median ratio 1.000, and no video off by more than one character. On those 55 videos Salin averaged 3.86s against 1.22s, because it pays for a full yt-dlp metadata extraction.

youtube-transcript-api drives the watch page directly and blocks after roughly 100 seconds of sequential requests (video #56 and #89 in two runs). Once blocked, it stayed blocked for the rest of the day.

Caption bytes come from the same api/timedtext endpoint for both libraries, so Salin shares that limit and can be rate-limited too. What differs is the response: Salin returns DownloadStatus.RATE_LIMITED or raises RateLimitedError with Retry-After, and keeps inspecting and resolving while the endpoint is closed. At ~11 videos/min the limit did not trip; Salin has not been shown to survive youtube-transcript-api-level throughput.

Other sites

Extraction goes through yt-dlp, so YouTube is not a hard limit, but there is no dedicated support beyond that. Another site works only if yt-dlp exposes its subtitle tracks in one of the supported formats.

Of the sites tried, only Dailymotion worked end to end (it offered srt). Vimeo returned an OAuth 401, TED raised an extractor error, and archive.org exposed no captions. Call inspect() first to see what a site actually offers.

Development

uv sync --all-groups
uv run ruff check
uv run ruff format --check
uv run ty check
uv run pytest

CI runs the same commands on Python 3.12 and 3.13.

License

MIT - see LICENSE.