Audio and Video Transcriber: text, SRT and VTT subtitles
Pricing
from $9.00 / 1,000 audio minute transcribeds
Audio and Video Transcriber: text, SRT and VTT subtitles
Transcribes audio and video files you link to or upload, and podcast RSS episodes, into text with timestamps, SRT and VTT subtitles and the detected language. A Whisper speech model runs inside the actor; no outside service is called.
Give it audio or video files and it returns what was said: the full transcript, timestamped segments, ready to use SRT and VTT subtitle files, and the spoken language it detected.
Speech recognition runs inside the actor with OpenAI's Whisper models (tiny, base and small, through faster-whisper). No outside speech service is called and no API key is needed. The models are released under the MIT License.
What you can send
- Direct file links: up to 1 GB each, in the formats under "Accepted files". From a video, the audio track is used.
- Uploaded files: upload from your computer in the input form. Choose a permanent or existing key-value store in the upload dialog, then pick that store under "Key-value stores with your files". The run gets read access to the stores you pick and nothing else.
- Key-value stores: every record in a picked store whose key ends in
.mp3,.wav,.flac,.ogg,.oga,.opus,.webm,.mkv,.mkaor.ogvis transcribed. - Podcast RSS feeds: the newest episodes of each feed, as many as you set.
Accepted files
- Containers: MP3, WAV, FLAC, Ogg (
.ogg,.oga,.ogv,.opus), WebM and MKV - Audio codecs: Opus, Vorbis, FLAC, MP3 and PCM
- Video codecs, when the file has a picture: AV1, VP8, VP9 and Theora
A file in any other format, or one that holds any other codec, gets an error row that names what it holds and lists what is accepted, and is not charged.
Links to pages on video and social platforms (YouTube, TikTok, Instagram, Facebook, X, Vimeo, Twitch, SoundCloud, Spotify and others) are refused with an error row. Send a direct link to a file you own or are allowed to process.
Options
- Speech model:
tiny(fastest),base(default),small(most accurate) - Spoken language: a two letter code such as
enores, or empty to detect it - Translate the speech to English
- Word timestamps: every word with its start, end and probability
- Names and terms to expect, so the model spells them right
- Skip silence and music (default on)
- Maximum minutes per file (default 120): only that many minutes from the start are transcribed and billed
Output
Dataset: one item per file with url, ok, language, languageProbability, durationSeconds, transcribedSeconds, speechSeconds, minutesCharged, wordCount, text, segments (each with start, end, text and, with word timestamps on, words), srt, vtt, and links to the .srt, .vtt and .txt files saved in the run's key-value store. Podcast episodes also carry title, published and feed. A file that cannot be downloaded or read gets ok: false and an error in plain words, and the run carries on with the next file.
Tested
- An English speech clip in FLAC, and the same clip as MP3 with cover art, as Opus in WebM with AV1 video, and as Opus in MKV with VP9 video
- An English Ogg Vorbis clip, and WAV and Opus music files
- A Spanish audiobook chapter in MP3 (language detected as
esat 0.973), and the same chapter translated to English with word timestamps - A NASA podcast RSS feed
- A file read from a picked key-value store
- Errors: a YouTube link, a web page, a PNG image, a missing file (HTTP 404), an empty file, a link that is not an RSS feed, MP4, M4A and AAC files, and MKV files with H.264 video or AAC audio
- The maximum charge: a file cut short at the limit and the next file left unprocessed
Speed
Apify gives a run one CPU core for every 4 GB of memory, so speed follows the memory you choose. With 2 GB and the base model, the first 2 minutes of a podcast episode, 101 seconds of it speech, took 48 seconds to transcribe.
Pricing
Pay per event: one event per started minute of audio transcribed. A 90 second file is two events. A file that fails is not charged. The run stops starting new files when it reaches the maximum charge you set, and a file is cut to the minutes that charge still covers.
Scope and conduct
Only the files, stores and feeds you supply are read. Files are deleted from the run's disk once decoded. The transcript is machine made and can contain mistakes. Speaker labels are not produced.
Support
Questions and problems: the Issues tab on this page. Replies are written by an AI assistant on behalf of the developer and reviewed by the developer.
