Language guide · Cantonese yue

Best local speech-to-text models for Cantonese

8 of the 26 reviewed local speech-to-text models explicitly list Cantonese. For long recordings on a mid-range NVIDIA GPU the matcher's first pick is Whisper turbo; none of them streams Cantonese natively, so live use relies on chunking runtimes; for meetings with speaker labels it is Whisper medium. 0 models offer native streaming, 0 native speaker diarization and 8 have a permissive license.

Top matches by setup

Computed with the same deterministic matcher as the homepage, restricted to models whose official source lists Cantonese. To keep the shortlist useful, each row shows the best-ranked model from three different families (for example, only one Whisper size). Scores are decision guidance, not accuracy claims.

Best-ranked model from each of the top three model families, Cantonese
Setup	#1	#2	#3
Long recordings, 8–12 GB NVIDIA GPU, balanced	Whisper turbo 79/100 · Strong fit	Qwen3-ASR 0.6B 75/100 · Strong fit	Fun-ASR-MLT-Nano-2512 75/100 · Strong fit
Live transcription (native streaming, speed first)	Qwen3-ASR 0.6B 77/100 · Strong fit	Whisper turbo 76/100 · Strong fit	Fun-ASR-MLT-Nano-2512 73/100 · Strong fit
Meetings with speaker labels (diarization + timestamps)	Whisper medium 84/100 · Strong fit	Qwen3-ASR 0.6B 77/100 · Strong fit	Fun-ASR-MLT-Nano-2512 72/100 · Strong fit
Subtitles with timing, accuracy first	Whisper medium 87/100 · Excellent fit	Qwen3-ASR 0.6B 80/100 · Strong fit	Fun-ASR-MLT-Nano-2512 68/100 · Possible fit
CPU-only machine, low memory use	Whisper tiny 74/100 · Strong fit	Qwen3-ASR 0.6B 58/100 · Possible fit	Fun-ASR-MLT-Nano-2512 58/100 · Possible fit

Re-rank for your setup

Change hardware, workload and requirements to re-rank all 8 Cantonese-capable models in your browser.

Choose by requirement

Live captions and voice input: No reviewed model streams Cantonese natively. Whisper-style models can be run through chunking runtimes; expect higher latency and test word splits at chunk borders.
Speaker-labelled meetings: No reviewed model that lists Cantonese documents native diarization; pair the transcription model with a dedicated diarization component.
Commercial products: Permissive licenses: Whisper (tiny, base, small, medium, turbo), Qwen3-ASR 0.6B, Qwen3-ASR 1.7B and Fun-ASR-MLT-Nano-2512. 0 other models carry attribution, ShareAlike or custom terms.
CPU-only or low-power machines: Rated 4/5 or better for CPU-only use: Whisper tiny, Whisper base and Whisper small.

All 8 models that list Cantonese

Reviewed local ASR models listing Cantonese
Model	Streaming	Timestamps	Diarization	License	CPU · NVIDIA fit
Fun-ASR-MLT-Nano-2512 FunAudioLLM · 800M	Via runtime	No	No	Apache-2.0 Permissive	2/5 · 5/5
Qwen3-ASR 0.6B Qwen · 0.6B	Via runtime	Separate aligner	No	Apache-2.0 Permissive	2/5 · 5/5
Qwen3-ASR 1.7B Qwen · 1.7B	Via runtime	Separate aligner	No	Apache-2.0 Permissive	1/5 · 5/5
Whisper base OpenAI · 74M	Via runtime	Yes	No	MIT Permissive	5/5 · 4/5
Whisper medium OpenAI · 769M	Via runtime	Yes	No	MIT Permissive	2/5 · 5/5
Whisper small OpenAI · 244M	Via runtime	Yes	No	MIT Permissive	4/5 · 5/5
Whisper tiny OpenAI · 39M	Via runtime	Yes	No	MIT Permissive	5/5 · 4/5
Whisper turbo OpenAI · 809M	Via runtime	Yes	No	MIT Permissive	2/5 · 5/5

Multilingual models that do not list Cantonese

Parakeet TDT 0.6B v3, Orukeet, Nemotron 3.5 ASR Streaming 0.6B, Omnilingual ASR CTC 300M, Cohere Transcribe 03-2026, VibeVoice-ASR 7B, VibeVoice-ASR Streaming 1.5B, VibeVoice-ASR-BitNet, Voxtral Mini 4B Realtime 2602, Granite Speech 5.0 2B, Granite Speech 5.0 470M TurboCTC, VibeVoice-ASR-Streaming-7B, Parakeet Redux, Whistle and Jais-ASR are multilingual but do not list Cantonese in the reviewed official source, so the matcher does not recommend them for this language. Omnilingual ASR documents 1,600+ language-script codes, so check Meta's full list ↗ before ruling it out.

Mandarin Chinese (zh) is a separate entry with far broader coverage; see the Mandarin Chinese guide. The Qwen3-ASR model cards additionally list 22 Chinese dialects.

Common questions

What is the best local speech-to-text model for Cantonese?

There is no single winner; it depends on the workload. In the matcher, Whisper turbo leads for long recordings on a mid-range GPU, Whisper medium for meetings with speaker labels and Whisper tiny for CPU-only machines. Benchmark the top two or three on your own Cantonese audio before deciding.

Is there an offline Cantonese model with real-time streaming?

Not natively among the reviewed models. You can stream through chunking runtimes, but latency and accuracy at chunk borders must be tested.

Does Whisper support Cantonese?

Yes. All Whisper sizes list Cantonese in the official tokenizer language list. Smaller sizes trade accuracy for speed; Whisper turbo is the usual starting point on a GPU.

Which models do not support Cantonese?

Among multilingual models, Parakeet TDT 0.6B v3, Orukeet, Nemotron 3.5 ASR Streaming 0.6B, Omnilingual ASR CTC 300M, Cohere Transcribe 03-2026, VibeVoice-ASR 7B, VibeVoice-ASR Streaming 1.5B, VibeVoice-ASR-BitNet, Voxtral Mini 4B Realtime 2602, Granite Speech 5.0 2B, Granite Speech 5.0 470M TurboCTC, VibeVoice-ASR-Streaming-7B, Parakeet Redux, Whistle and Jais-ASR do not list Cantonese in their reviewed sources. English-only models are also excluded.

Generated from the reviewed model database (latest review November 3, 2026). "Lists Cantonese" means the model's official source explicitly names the language; it does not guarantee equal accuracy across languages.