Whisper Small Cantonese: a community Cantonese fine-tune
Don’t want to install Whisper Small Cantonese? Drop the file here.
The first 20 minutes of any recording are transcribed free: a recording up to 20 minutes transcribes in full, a longer one is previewed, and a pack from $6 (500 minutes) transcribes the rest and opens every export, and credits never expire.
Whisper Small retrained on about 934 hours of Cantonese, from Common Voice to pseudo-labelled YouTube audio, scoring 7.93 CER on Common Voice 16 Cantonese without punctuation.
Get Whisper Small Cantonese Official site & pricing
Whisper Small Cantonese (alvanlii) is whisper-small retrained on ~934 hours of Cantonese; 7.93 CER on Common Voice 16 without punctuation; ~1.5 GB of VRAM. Apache-2.0.
A good base for a Cantonese pipeline you run yourself. Compare SenseVoice and CUHK's work; or skip the hosting and upload the recording.
What it is
alvanlii/whisper-small-cantonese is a community model on Hugging Face: OpenAI's whisper-small fine-tuned for Cantonese and released under Apache-2.0. The model card lists its training data: Common Voice 16.0 zh-HK (138 h) and yue (85 h), Common Voice 17.0 yue (178 h), Cantonese-ASR (72 h), CantoMap (23 h) and 438 hours of pseudo-labelled YouTube audio, about 934 hours in all. On the Common Voice 16.0 Cantonese test set it reports a character error rate of 9.72 with punctuation and 7.93 without, down from 10.73 and 15.81 in earlier versions. On an RTX 3090 the card measures 0.055 s per sample with fast attention and about 1.5 GB of VRAM; on CPU, 2.57 s per sample.
Watch out for: A single community fine-tune of the small Whisper size: 9.72 CER with punctuation on Common Voice, which is read speech; expect more errors on noisy, fast or code-switched conversation. You host it.
Cantonese recordings, without hosting a model
- Drop the audio or video file into the box at the top of this page, or paste a link (YouTube, a podcast or RSS feed, Google Drive, Twitch or Kick). Nothing to install and no GPU needed.
- Cantonese is among the languages detected automatically, and the transcript stays in the language that was spoken. WhipScribe does not translate.
- Read the summary free. A pack from $6 (500 minutes) opens the full transcript and the downloads: TXT, DOCX, SRT, VTT and JSON. Packs are one-time purchases and credits never expire.
- Files can be up to 10 hours or 5 GB per file.
Related: Transcribe Cantonese · SenseVoice · CUHK Cantonese ASR · SRT generator · Pricing
Install / use
pip install transformers torch librosa # model: alvanlii/whisper-small-cantoneseKey specs
What people actually do with Whisper Small Cantonese-style transcription
The tool is the means. These are the jobs — each one priced at published rates, each one wired up on its own page.
Features
| Speaker diarization | Not stated |
| Word-level timestamps | Not stated |
| Streaming / real-time | Not stated |
| Languages supported | Cantonese |
| HIPAA eligible | No |
Links
- alvanlii/whisper-small-cantonese ↗model card: data, CER, speed · Apache-2.0
- Cantonese YouTube pseudo-transcription ↗the 438-hour pseudo-labelled set
Whisper Small Cantonese vs Whipscribe
| Feature | Whisper Small Cantonese | Whipscribe |
|---|---|---|
| Category | Open source | Transcription APIs |
| Pricing | free (Apache-2.0) | $6–$24 one-time packs (credits never expire) · first 20 minutes free |
| Speaker identification | No | Named in the summary, key quotes and chapters |
| Word timestamps | No | Yes |
| Streaming | No | No |
| Languages | Cantonese | 99 |
| Platforms | Linux, macOS, Windows | Web, API, MCP |
Alternatives to Whisper Small Cantonese
Frequently asked about Whisper Small Cantonese
How accurate is Whisper Small Cantonese?
On the Common Voice 16.0 Cantonese test set the model card reports 9.72 character error rate with punctuation and 7.93 without. Common Voice is read speech, so conversational audio will score worse.
What was it trained on?
About 934 hours: Common Voice 16.0 zh-HK (138 h) and yue (85 h), Common Voice 17.0 yue (178 h), Cantonese-ASR (72 h), CantoMap (23 h) and 438 h of pseudo-labelled YouTube audio.
What hardware does it need?
Very little for a Whisper model: the card measures about 1.5 GB of VRAM and 0.055 s per sample on an RTX 3090 with fast attention, or 2.57 s per sample on CPU.
Is Whisper Small Cantonese free for commercial use?
It is released under Apache-2.0. Check the licences of its training datasets if that matters for your use.
Can WhipScribe transcribe Cantonese?
Yes. Cantonese is among the languages WhipScribe detects automatically, and the transcript stays in the language that was spoken. WhipScribe does not use this model.
What is Whisper Small Cantonese?
Whisper Small Cantonese (alvanlii) is whisper-small retrained on ~934 hours of Cantonese; 7.93 CER on Common Voice 16 without punctuation; ~1.5 GB of VRAM. Apache-2.0.
Does Whisper Small Cantonese work on Mac, Windows and Linux?
Whisper Small Cantonese runs on Linux, macOS and Windows.
How do I install Whisper Small Cantonese?
With pip: `pip install transformers torch librosa # model: alvanlii/whisper-small-cantonese`. You need Python and, for most audio tools, ffmpeg on your PATH first.
Is Whisper Small Cantonese HIPAA compliant?
Whisper Small Cantonese does not claim HIPAA compliance. For PHI, run it on infrastructure you control or choose a service that will sign a BAA.
What are the limitations of Whisper Small Cantonese?
A single community fine-tune of the small Whisper size: 9.72 CER with punctuation on Common Voice, which is read speech; expect more errors on noisy, fast or code-switched conversation. You host it.
Who is Whisper Small Cantonese best for?
Developers who need a small, cheap-to-run Cantonese model they can fine-tune further or run on one consumer GPU.
What should I know before choosing Whisper Small Cantonese?
A good base for a Cantonese pipeline you run yourself. Compare SenseVoice and CUHK's work; or skip the hosting and upload the recording.
Does Whisper Small Cantonese run offline?
Whisper Small Cantonese runs on your own machine — it is open source, so nothing leaves the computer unless you configure it to.
Can I use Whisper Small Cantonese in the browser?
No — Whisper Small Cantonese is a Linux, macOS and Windows application, not a web app. You install it rather than sign in to it.
Who makes Whisper Small Cantonese?
Whisper Small Cantonese is made by alvanlii.
Whipscribe is a hosted transcription service. If you want transcripts without running infrastructure, paste a URL or drop a file in the form below — you'll have a transcript in seconds.
Transcribe your recordings
Running Whisper Small Cantonese yourself? Send the same file here and compare the transcript. Transcribe any language, any recording · Transcription services from $0.29 an hour · Transcribe from any meeting or recording platform · How to transcribe, step by step
Explore
All transcription tools · By language · By platform · By file format · How-to guides · Transcription services · Transcription API · Pricing
Long recording to get through?
One file up to 10 hours, as an upload or a link: see MP3 to text for long recordings, transcribing a podcast episode and the SRT file generator.
Rather have it done for you?
Free tools cost setup time. When you want a finished transcript instead: see the best transcription services side by side, the cheapest way to transcribe an hour of audio and transcription services from $6.