Transcribe audio and video online

From an audio file, video or YouTube link

Start with the recording you want to read, quote or caption. You can upload an interview, meeting, podcast or lecture in a supported format such as MP3, WAV, M4A, MP4 or MOV. With a video file, Transcriber extracts the soundtrack and recognizes the speech on that track. Any lettering already visible in the picture, such as a slide title or a lower third, stays outside the transcript unless someone also says it aloud.

Local uploads support up to 512 MiB and 30 minutes per recording. For longer supported YouTube videos, Transcriber uses the first 30 minutes. If your interview begins 45 minutes into a panel recording, prepare the relevant excerpt from a copy you have permission to process. A transcript of that excerpt starts at zero. Keep a note of the original offset if someone will check your citations against the full recording.

The original recording is often the best starting point. If you have a separate microphone track and a room recording, listen to the same short exchange in both before uploading. Pick the version where you can distinguish the speakers and hear the ends of their sentences. A loud track can still be hard to transcribe if music masks the speech or if the microphone distorted the loudest words.

Prepare a clip with Media Converter →

How Whisper turns the soundtrack into text

CS Match prepares the audio for its transcription worker, which runs the open-source Whisper model through faster-whisper and CTranslate2. Whisper converts acoustic input into text tokens. It uses the sound and the context of the spoken passage to choose a sequence of words, which is why it can produce readable punctuation and phrasing from a continuous recording. The worker returns text in segments with time ranges, so the same result can drive playback, document export and subtitle cues.

The recording language tells Whisper which language to recognize. Choose it explicitly when you know what most of the recording contains, especially when a short greeting or introduction is in another language. Auto detect is useful when you are unsure. An interview that switches languages halfway through can still require review in both languages. The transcription task is designed to write down the spoken language. If you need a translated version, treat translation as a separate step.

Voice activity detection helps distinguish speech from stretches of silence and background sound. The difficult cases are often quiet speech under music, voices far from a microphone, sudden edits and several people talking at once. Names, product terminology and numbers deserve a listen even when a sentence looks fluent. For example, a believable company name in the transcript might be a different company with a similar-sounding name. The model can suggest a plausible spelling, but the recording and your source material are what let you verify it.

Whisper model documentation →faster-whisper documentation →

Choose between Quick pass, Balanced and Careful pass

These modes change how many candidate text sequences the decoder can consider at each step. Quick pass uses a beam size of one, Balanced uses three and Careful pass uses five. A larger beam gives Whisper more alternatives to compare and can take longer. All three modes use the same underlying model in this workflow. Recording quality and language matter enough that a wider search can still choose the wrong word.

Why a YouTube video can play but fail to import

A YouTube watch URL identifies a page. The browser player fetches media separately, with its own playback context. Google’s YouTube Data API can return details such as the title, duration and availability information. To transcribe, our server also needs the audio bytes. A video can open in your signed-in browser while a server request for the same video gets a different response. Access restrictions, removed videos and temporary errors at the media provider can all prevent a usable audio file from being prepared.

When you paste a supported public link, CS Match sends an import request to a media provider. The provider prepares an MP3 and returns a temporary file URL when it is ready. CS Match downloads the file, checks the audio and trims longer recordings to the first 30 minutes before handing them to Whisper. This can involve transferring more than 30 minutes of audio before the cut. The duration limit describes what is transcribed. If the provider cannot prepare or deliver the MP3, the task fails before speech recognition begins, so switching from Balanced to Careful pass cannot resolve it.

Use the link to a single available video. If an import fails while YouTube still plays the video for you, try again later or upload a copy you have permission to use. For a video you published, you can use your original file or download your own upload through YouTube Studio where available. Uploading a file goes straight to media preparation, so it also gives you a way to test whether a problem belongs to the link or the audio itself.

YouTube video metadata reference →Download your own YouTube uploads →

Review the transcript with speakers and playback

For a conversation, turn on speaker detection and choose Auto or specify a count from one to five. This adds a separate analysis that groups voice segments by acoustic similarity and associates them with timed text. Labels such as Speaker 1 and Speaker 2 help you follow the turns in that recording. A host can receive a second label after a microphone change, and two guests with similar voices can be grouped together. A count you already know is useful input, especially in a two-person interview with a consistent setup.

Open a completed transcript from Recent files. The audio player sits at the bottom of the transcript view, and the current passage highlights during playback. Enable Show timestamps and select a time to jump back to that point. If the wording of a quote matters, listen from a few seconds before the highlighted segment through the end of the next sentence. The surrounding audio often clarifies whether a speaker was agreeing, repeating someone else or correcting a number.

For a meeting, start by checking the people, owners and dates mentioned around each decision. For an interview, verify both the words and which person said them. If someone speaks off-mic and the transcript drops the line entirely, add a note in your exported document after listening. The best review method depends on what the text will be used for. An internal outline can tolerate a rough passage that would need correction in a published quotation or a caption track.

Learn how speaker turns are grouped →

Download text, documents or subtitles

Select Export transcript in Recent files to download a finished job, or open the transcript first and listen through the passages you need. With PDF, DOCX and TXT you can include timestamps when timed segments are available. CSV stores timing in columns, and SRT and VTT use the start and end of each cue to control when it appears on screen. Speaker labels can be included where the recording produced them. Plain TXT without timestamps gives you the continuous text.

Use DOCX when you still need to correct spellings, add headings or leave comments for an editor. Use PDF when the text is ready for someone to read or archive. For a spreadsheet, choose CSV. Separate start, end, speaker and text fields make it easier to sort or review individual passages. Keep both start and end times if a colleague will use your selection to cut clips from the recording.

Explore the subtitle tools →WebVTT specification →

Four ways to use a finished transcript

For a recorded interview, run Balanced with speaker detection set to two if you know it contains one interviewer and one guest. Review the turns at each interruption, then export a DOCX with timestamps for the editor. Keep the timestamps in your working copy so a reviewer can hear the original context without searching through the entire recording.

For a training video, export SRT or VTT and load the file in the player or editor where you will publish it. Play back the start, middle and end to check synchronization. Shorten lines that disappear too quickly to read, and add important non-speech cues by hand if the audience needs them. If you edit the video after generating captions, check the cue times again against the new cut.

For meeting notes, export CSV when you want to tag passages by topic or extract the exact times of decisions. Choose TXT when you only need a searchable record. Before circulating a decision or assigning a task from the transcript, verify the action, the owner and any deadline. Those details are easy to mishear when several people speak in quick succession.

For a long podcast episode, split a permitted source file at a natural break, such as a chapter boundary or the end of a guest’s answer. Transcribe each part with the same language and mode. Record where each part begins in the original episode, since its exported timestamps start from the new file. When you assemble notes across parts, treat each part’s speaker labels separately. Speaker 1 in the second file can refer to a different person than Speaker 1 in the first.