Local ASR models
4 local speech recognition models for Windows compared
Compare the complete language coverage and user-facing features of SenseVoiceSmall, Parakeet, FireRedASR2-AED, and Qwen3-ASR-0.6B.
How the four local ASR models differ
Owl Meeting converts all four speech recognition models from their original versions into CPU-friendly ONNX format and optimizes them for local Windows use. No dedicated GPU is required: a regular PC can use its CPU for fast live recognition or offline transcription, while keeping core speech recognition and recordings on the device.
| Model | Main strength | Language focus | Local package | Best starting point |
|---|---|---|---|---|
| SenseVoiceSmall | Speed | Mandarin, Cantonese, English, Japanese, and Korean | 230 MB quantized / 900 MB standard | Live captions and low-latency meetings |
| Parakeet TDT 0.6B v3 | Speed + multilingual | 25 European languages with automatic language identification | About 850 MB | Real-time European multilingual recognition |
| FireRedASR2-AED | Chinese accuracy | Mandarin, Chinese dialects and accents, English, and code-switching | About 1.6 GB; punctuation model adds about 390 MB | High-accuracy offline Chinese and dialect transcription |
| Qwen3-ASR-0.6B | Accuracy + multilingual | 30 languages and 22 Chinese dialects | About 895 MB | High-accuracy offline multilingual, dialect, and difficult audio |
SenseVoiceSmall
Speed
Owl Meeting · Model 1
SenseVoiceSmall uses a non-autoregressive architecture for low-latency transcription in Mandarin, Cantonese, English, Japanese, and Korean.
- Automatic language identification removes the need to select a language for every recording.
- It can label emotion and detect BGM, applause, laughter, crying, coughing, and sneezing.
- Owl Meeting offers an approximately 230 MB quantized package and a 900 MB standard package.
Supported languages and dialects
Mandarin (zh), Cantonese (yue), English (en), Japanese (ja), Korean (ko)
Parakeet TDT 0.6B v3
Speed + multilingual
Owl Meeting · Model 3
Parakeet TDT 0.6B v3 is a 600-million-parameter FastConformer-TDT model for responsive European multilingual meetings, interviews, and subtitles.
- It covers 25 European languages and identifies the spoken language automatically without prompting.
- Output can include punctuation, capitalization, word-level timestamps, and segment-level timestamps.
- Long-audio transcription suits meetings, interviews, classes, and subtitle production.
Supported languages and dialects
Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru), Ukrainian (uk)
FireRedASR2-AED
Chinese accuracy
Owl Meeting · Model 2
FireRedASR2-AED covers Mandarin, more than 20 Chinese dialects and accents, English, and Chinese-English code-switching. Owl Meeting uses it for high-accuracy offline transcription.
- Coverage includes Cantonese, Sichuan, Shanghai, Wu, Minnan, and many regional accents.
- It supports both speech and singing; the AED variant also exposes word timestamps and confidence scores.
- The companion FireRedPunc model adds Chinese and English punctuation; Owl Meeting packages it separately.
Supported languages and dialects
Mandarin (zh), English (en)
Chinese dialects and accents: Cantonese (Hong Kong and Guangdong), Sichuan, Shanghai, Wu, Minnan, Anhui, Fujian, Gansu, Guizhou, Hebei, Henan, Hubei, Hunan, Jiangxi, Liaoning, Ningxia, Shaanxi, Shanxi, Shandong, Tianjin, Yunnan, and others
Qwen3-ASR-0.6B
Accuracy + multilingual
Owl Meeting · Model 4
Qwen3-ASR-0.6B covers 30 languages and 22 Chinese dialects for multilingual, dialect, and difficult-audio offline transcription.
- One model provides automatic language identification plus both offline and streaming inference.
- It handles speech, singing voice, and songs with background music, with long-audio support.
- For precise timestamps, pair it with the separate Qwen3-ForcedAligner-0.6B.
Supported languages and dialects
Chinese (zh), English (en), Cantonese (yue), Arabic (ar), German (de), French (fr), Spanish (es), Portuguese (pt), Indonesian (id), Italian (it), Korean (ko), Russian (ru), Thai (th), Vietnamese (vi), Japanese (ja), Turkish (tr), Hindi (hi), Malay (ms), Dutch (nl), Swedish (sv), Danish (da), Finnish (fi), Polish (pl), Czech (cs), Filipino (fil), Persian (fa), Greek (el), Hungarian (hu), Macedonian (mk), Romanian (ro)
Chinese dialects and accents: Anhui, Dongbei, Fujian, Gansu, Guizhou, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningxia, Shandong, Shaanxi, Shanxi, Sichuan, Tianjin, Yunnan, Zhejiang, Cantonese (Hong Kong accent), Cantonese (Guangdong accent), Wu, Minnan
Real-time use: SenseVoice and Parakeet
Live recognition depends on sustained latency and partial-result stability, not only final WER. SenseVoice is the speed-first option for Chinese, English, Japanese, or Korean meetings. Parakeet is the stronger starting point for European multilingual meetings that still need a responsive live experience.
High-accuracy offline use: FireRedASR and Qwen3
FireRedASR2-AED targets Mandarin, dialects, accents, and Chinese-English code-switching. It records 3.05% average CER across four public Mandarin benchmarks and 11.67% across 19 Chinese dialect and accent benchmarks.
Qwen3-ASR-0.6B supports 30 languages and 22 Chinese dialects with both offline and streaming inference for multilingual meetings, dialect recordings, and difficult audio.
A practical starting choice
- Real-time, Chinese/English/Japanese/Korean, speed first: try SenseVoiceSmall.
- Real-time, English and European multilingual: try Parakeet TDT 0.6B v3.
- Offline, Chinese and dialect accuracy first: try FireRedASR2-AED.
- Offline, accuracy plus broad multilingual coverage: try Qwen3-ASR-0.6B.