Speech Recognition (Video): one video in, one transcript out, and where to run it

OpenAI released Whisper Large V3 under Apache 2.0, so the weights are free and commercial use is unrestricted. At 1.55 billion parameters, self-hosting means the Hugging Face transformers library and a GPU for speed, though it runs on a CPU if you are patient. In your browser it runs on cnaps studio, included in the basic plan.

What this particular setup does is remove a chore. Transcribing a video normally means extracting the audio track first, then feeding that to a speech model. Here you hand over the video file and get back the text, with the audio extraction handled for you, which is the difference between a two-step job and a one-step one.

The cost is that it gives you one flat block of text with no timing information. If you need to know when each phrase was said, so captions line up, that is the sibling in this catalog rather than this one. Two speech recognition entries sit in the cnaps studio catalog and this is the simpler of the two.

What it actually is

An encoder-decoder transformer trained on five million hours of audio: one million hours weakly labelled and four million pseudo-labelled, run for two passes over the data. That scale is the reason it copes with accents, background noise and technical vocabulary better than models trained on clean studio recordings.

The model works in 30-second windows and stitches longer audio together behind the scenes, so a two-hour recording needs no chunking from you. It also carries several safeguards against the failure mode these models are known for, where a stretch of silence produces invented text: thresholds on compression ratio, on log probability, and on the model's own no-speech signal, with a fallback that retries a difficult segment.

Two settings matter. Spoken language defaults to Auto, which detects the language from the audio, with explicit choices for English, Korean, Japanese and Chinese; the model itself covers 99 languages, so Auto usually works for anything else. Output mode is either Transcribe, which keeps the spoken language, or Translate to English, which turns any language into English text directly. Running both gives you a bilingual pair from one recording.

Try Speech Recognition (Video) in cnaps studio →

{
  "text": "Welcome back. Today, we'll turn one product photo into a short ad video. No editing software, just a single flow in CNAP's studio.",
  "words": [
    {"text": "Welcome", "start_sec": 0.0, "end_sec": 0.3},
    ...
    {"text": "CNAP's", "start_sec": 7.84, "end_sec": 8.4},
    {"text": "studio.", "start_sec": 8.4, "end_sec": 8.8}
  ]
}

{
  "text": "Welcome back. Today we'll turn one product photo into a short ad video. No editing software, just a single flow in CNAPS Studio.",
  "words": [
    ...
    {"text": "CNAPS", "start_sec": 7.84, "end_sec": 8.4},
    {"text": "Studio.", "start_sec": 8.4, "end_sec": 8.8}
  ],
  "script_alignment": {"applied": true, "ratio": 0.9565, "matched": 22}
}

The published numbers

Two figures come from the Open ASR Leaderboard, which is a third-party ranking rather than a vendor claim, and that makes this one of the better-evidenced models in the whole series.

The accuracy figure is a mean word error rate of 7.44. Word error rate counts mistakes rather than successes, so lower is better, and 7.44 means roughly seven or eight words in every hundred come back wrong, counting substitutions, insertions and deletions across the leaderboard's mix of datasets. For dictated or well-recorded speech you should expect better than that average; for a noisy room with crosstalk, worse.

The speed figure is a real-time factor of 145.51, meaning it processes about 145 seconds of audio per second of compute on capable hardware. An hour-long recording is therefore a matter of tens of seconds rather than an hour, which is what makes transcribing a back catalogue practical.

Both numbers come from an independent leaderboard that re-runs models under common conditions, which is a genuinely different class of evidence from the vendor-reported figures elsewhere in this series. Treat them as reliable, and treat your own audio as the thing that shifts them.

How it compares to the other speech model here

The accuracy row is identical because these are the same model behind two different input and output shapes. Nothing about transcription quality separates them, so a comparison of scores would be meaningless.

Choose by what you plan to do with the words. Reach for Speech Recognition (Video) when you want readable text: a meeting record, an interview to quote from, show notes, or something to search. Reach for Speech Recognition (Clips) when the text has to line up with pictures, which is subtitles and captions. If the audio is under thirty seconds and you also want it summarised, Gemma 4 E2B does both in one step.

What to chain it with

A video goes in and text comes out, which puts this at the boundary between the video half of the catalog and the language half.

Two chains do the real work. The obvious one is understanding: send the transcript to Gemma 4 31B or Qwen3.6-35B-A3B and ask for a summary, a list of decisions, or an answer to a specific question, which turns an hour of recording into something you read in a minute. The other is translation without a translation step, using the Translate to English mode to go straight from foreign-language audio to English text, then on to a language model for a cleaned-up version.

Open Speech Recognition (Video) in cnaps studio, transcribe your worst recording rather than your best one, and see where the 7.44 average sits against your actual material.

Start for free →