Speech Recognition (Clips): timed transcripts for a whole list of clips, and where to run it

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

What separates this from its sibling is the shape of the job. It takes a list of clips rather than one video, and returns a timed transcript for each, in a single pass. Split a recording into scenes or highlights upstream and you are left with a pile of segments that each need captioning; this handles the pile rather than making you wire up one transcription step per clip.

The timestamps are the other half of the point. Because each transcript carries timing, the text can be lined up against the original timeline for subtitles or search markers rather than sitting as a loose block of prose. Two speech recognition entries sit in the CNAPS Studio catalog and this is the one built for video work.

What it actually is

The same model as its sibling, wrapped for a different input. An encoder-decoder transformer trained on five million hours of audio, one million weakly labelled and four million pseudo-labelled, over two passes. Nothing about accuracy differs between the two entries; only what goes in and what comes out.

The model reads audio in 30-second windows and stitches longer stretches together itself, and it carries several guards against inventing words during silence: thresholds on compression ratio, on log probability and on its own no-speech signal, plus a retry for difficult segments. Timestamps come from the model's own timing prediction rather than a separate alignment step.

Two settings, the same as the sibling. Spoken language defaults to Auto, which detects per clip, so a mixed-language list is handled naturally without splitting it; explicit choices cover English, Korean, Japanese and Chinese, and the model itself knows 99. Output mode is Transcribe to keep the spoken language, or Translate to English. Running the list twice in both modes gives you a bilingual subtitle stack from one pass of work.

Base modelWhisper Large V3Parameters1.55 billionInputA list of video clips, typically from a splitter upstreamOutputOne timed transcript per clipTimestamp granularitySentence level or word levelLanguage detectionPer clip, so mixed-language lists workLanguages99 in the model; Auto, English, Korean, Japanese and Chinese in the pickerOutput modesTranscribe, or translate to EnglishTraining data5 million hours of audio, two passesLicenseApache 2.0

The published numbers

The accuracy figures come from the Open ASR Leaderboard, an independent ranking that re-runs models under common conditions, which makes this better evidenced than most models in this series.

The headline is a mean word error rate of 7.44, and word error rate counts mistakes rather than successes, so lower is better. Roughly seven or eight words in a hundred come back wrong across the leaderboard's mixture of datasets, counting words swapped, added and dropped. Clean dictation does better than that; a noisy room with people talking over each other does worse.

Speed is a real-time factor of 145.51, meaning about 145 seconds of audio processed per second of compute on capable hardware. For a list of clips that matters more than for a single video, because the cost is per second of audio regardless of how many pieces it arrives in.

One thing is not published: nothing measures timestamp accuracy specifically. The word error rate tells you how well it hears, not how precisely the timings land, and for subtitle work the second question is the one that decides whether captions feel right. Check that on your own footage.

How it compares to the other speech model here

Speech Recognition (Clips)Speech Recognition (Video)InputA list of video clipsOne video fileOutputA timed transcript per clipOne plain-text transcriptTimestampsYes, sentence or word levelNoBest forSubtitles, per-scene captions, searchable segmentsMeetings, lectures, interviews, show notesUnderlying modelWhisper Large V3Whisper Large V3Word error rate7.44 mean7.44 meanLicenseApache 2.0Apache 2.0

The word error rate is identical because it is the same model. Anyone presenting these two as an accuracy choice would be misreading the table; the choice is entirely about input and output shape.

Reach for Speech Recognition (Clips) when something has already cut your video into pieces and each piece needs its own text, or when the text has to align with pictures. Reach for Speech Recognition (Video) when you have one recording and want one readable transcript. For audio under thirty seconds where you also want the content interpreted rather than just transcribed, Gemma 4 E2B takes audio directly and answers questions about it.

What to chain it with

A list of clips goes in and a list of timed transcripts comes out, so this sits in the middle of a video pipeline rather than at either end.

Two chains use it properly. For subtitling, Video Split cuts the source into segments, this transcribes them with timing, and the timed text drives captions back onto the footage. For a searchable archive, pair it with Subject Follow (9:16) or a highlight step so each clip carries both a vertical cut and the words spoken in it, which is what makes a back catalogue findable by phrase rather than by filename.

Either way, a language model such as Gemma 4 31B reading the collected transcripts turns a folder of clips into a summary of what is in them.

Open Speech Recognition (Clips) in CNAPS Studio, run a handful of clips through it, and check the timestamps against the footage before you build a subtitle pipeline on top.

Sources

• huggingface.co/openai/whisper-large-v3 for the Apache 2.0 license, the 1.55 billion parameter count, the architecture, the 5 million hours of training audio, the 99 languages, the 30-second window, the timestamp prediction, and the anti-hallucination thresholds.
• arXiv:2212.04356 for the training method behind the model.
• The Open ASR Leaderboard for the 7.44 mean word error rate and the 145.51 real-time factor.
• docs.cnaps.ai/ai-model-hub for the CNAPS Studio input and output shapes, the per-clip language detection, the language picker and the output modes.