Speech-to-text models for real-time and file transcription-Alibaba Cloud Model Studio(Model Studio)-阿里云帮助中心

Use the four "Selection criteria"—real-time vs. offline, domain terminology, speaker diarization, and emotion recognition—to identify the right model. Then check "Recommended models" for top picks per scenario, "All models" for the full version list by model family, and "Audio specifications" for input constraints. Supported languages (including dialects) are listed under each model family in "All models".

Selection criteria

Evaluate the following four dimensions to narrow down the right model for your use case.

Real-time or offline?

Real-time recognition outputs results while the user is still speaking. Offline recognition transcribes a pre-recorded audio file after the recording ends.

Real-time (streaming recognition): Uses a WebSocket connection to stream audio in and text out. Ideal for live captions, voice assistants, and meeting transcription. Recommended models: fun-asr-realtime (hot words, dialect support) or qwen3.5-omni-plus-realtime (prompt context, multilingual).
Offline (file transcription): Uses an HTTP API to submit audio files and retrieve transcription results. Suited for call center recordings, podcasts, and interviews. Recommended models: fun-asr (hot words, speaker diarization) or qwen3.5-omni-plus (prompt context, multilingual).

Fun-ASR and Qwen-ASR real-time models support the DashScope SDK (Java, Python). Fun-ASR models also support Android and iOS SDKs. All other models require direct WebSocket or HTTP API calls.

For real-time integration, see Real-time speech recognition. For offline integration, see Non-real-time speech recognition.

Domain terminology

Two approaches, ranked by flexibility:

Prompt context injection: Describe your domain in the system prompt. No setup required—the model adapts on each request. Trade-off: slightly higher per-request latency than dedicated ASR models. Use qwen3.5-omni-plus-realtime (real-time) or qwen3.5-omni-plus (offline).
Hot words: Supply a weighted vocabulary list. Best for stable, infrequently changing terminology. Use fun-asr-realtime (real-time) or fun-asr (offline).

Note

Qwen3.5-Omni isn't a traditional ASR engine—it's a large language model that understands audio. Inject context through the system prompt; the model adapts without a hot word list.

Speaker diarization

Only the Fun-ASR offline models (fun-asr and fun-asr-mtl) support speaker diarization. Use these models when you need to identify "who said what."

Emotion recognition

Qwen-ASR and Qwen3.5-Omni models detect emotions alongside transcription. Recommended models: qwen3-asr-flash-realtime (real-time) or qwen3-asr-flash-filetrans (offline).

Recommended models

The following table lists top picks for common scenarios. For full details, visit Model Studio.

Model ID	Mode	API	Accuracy boost	Emotion recognition	Speaker diarization	Languages	Max audio duration/size
`fun-asr-realtime`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese, English, Japanese, and dialects	Unlimited
`fun-asr`	Offline	HTTP	Hot words	Unsupported	Supported	Multilingual with dialects	12 hours / 2 GB
`qwen3.5-omni-plus-realtime`	Real-time	WebSocket	Prompt context	Supported	Unsupported	Multilingual	2 hours
`qwen3.5-omni-plus`	Offline	HTTP (OpenAI-compatible)	Prompt context	Supported	Unsupported	Multilingual	3 hours / 2 GB

All models

Fun-ASR

Model ID	Mode	API	Accuracy boost	Emotion recognition	Speaker diarization	Languages	Max audio duration/size
`fun-asr-realtime`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese, English, Japanese, and dialects	Unlimited
`fun-asr-realtime-2026-02-28`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese, English, Japanese, and dialects	Unlimited
`fun-asr-realtime-2025-11-07`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese, English, Japanese, and dialects	Unlimited
`fun-asr-realtime-2025-09-15`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese, English	Unlimited
`fun-asr-flash-8k-realtime`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese	Unlimited
`fun-asr-flash-8k-realtime-2026-01-28`	Real-time	WebSocket	Hot words	Unsupported	Unsupported	Chinese	Unlimited
`fun-asr`	Offline	HTTP	Hot words	Unsupported	Supported	Multilingual with dialects	12 hours / 2 GB
`fun-asr-flash-2026-06-15`	Offline	HTTP	Prompt context	Unsupported	Unsupported	Multilingual with dialects	5 minutes / 2 GB
`fun-asr-2025-11-07`	Offline	HTTP	Hot words	Unsupported	Supported	Multilingual with dialects	12 hours / 2 GB
`fun-asr-2025-08-25`	Offline	HTTP	Hot words	Unsupported	Supported	Chinese, English	12 hours / 2 GB
`fun-asr-mtl`	Offline	HTTP	Hot words	Unsupported	Supported	Multilingual with dialects	12 hours / 2 GB
`fun-asr-mtl-2025-08-25`	Offline	HTTP	Hot words	Unsupported	Supported	Multilingual with dialects	12 hours / 2 GB

Supported languages (by version):

Fun-ASR-Realtime main versions (fun-asr-realtime, fun-asr-realtime-2026-02-28, fun-asr-realtime-2025-11-07): Chinese (Mandarin, Cantonese, Wu, Hokkien, Hakka, Gan, Xiang, Jin; plus regional accents including Central Plains, Southwestern, Ji-Lu, Jianghuai, Lan-Yin, Jiao-Liao, Northeastern, Beijing, and Hong Kong/Taiwan—covering Henan, Shaanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, Hebei, Tianjin, Shandong, Anhui, Nanjing, Jiangsu, Hangzhou, Gansu, and Ningxia accents), English, Japanese
fun-asr-realtime-2025-09-15: Chinese (Mandarin), English
Fun-ASR-Flash-8K-Realtime (fun-asr-flash-8k-realtime, fun-asr-flash-8k-realtime-2026-01-28): Chinese
Fun-ASR / Fun-ASR-MTL main versions (fun-asr, fun-asr-2025-11-07): Chinese (Mandarin, Cantonese, Wu, Hokkien, Hakka, Gan, Xiang, Jin; plus regional accents including Central Plains, Southwestern, Ji-Lu, Jianghuai, Lan-Yin, Jiao-Liao, Northeastern, Beijing, and Hong Kong/Taiwan—covering Henan, Shaanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, Hebei, Tianjin, Shandong, Anhui, Nanjing, Jiangsu, Hangzhou, Gansu, and Ningxia accents), English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, Slovak
Fun-ASR-Flash (fun-asr-flash-2026-06-15): Chinese (Mandarin, Cantonese, Wu, Hokkien, Hakka, Gan, Xiang, Jin; plus regional accents including Central Plains, Southwestern, Ji-Lu, Jianghuai, Lan-Yin, Jiao-Liao, Northeastern, Beijing, and Hong Kong/Taiwan—covering Henan, Shaanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, Hebei, Tianjin, Shandong, Anhui, Nanjing, Jiangsu, Hangzhou, Gansu, and Ningxia accents), English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, Slovak
fun-asr-2025-08-25: Chinese (Mandarin), English
Fun-ASR-MTL (fun-asr-mtl, fun-asr-mtl-2025-08-25): Chinese (Mandarin, Cantonese), English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, Slovak

Qwen-ASR

Model ID	Mode	API	Accuracy boost	Emotion recognition	Speaker diarization	Languages	Max audio duration/size
`qwen3-asr-flash-realtime`	Real-time	WebSocket	Unsupported	Supported	Unsupported	Multilingual with dialects	Unlimited
`qwen3-asr-flash-realtime-2026-02-10`	Real-time	WebSocket	Unsupported	Supported	Unsupported	Multilingual with dialects	Unlimited
`qwen3-asr-flash-realtime-2025-10-27`	Real-time	WebSocket	Unsupported	Supported	Unsupported	Multilingual with dialects	Unlimited
`qwen3-asr-flash-filetrans`	Offline	HTTP	Unsupported	Supported	Unsupported	Multilingual with dialects	12 hours / 2 GB
`qwen3-asr-flash-filetrans-2025-11-17`	Offline	HTTP	Unsupported	Supported	Unsupported	Multilingual with dialects	12 hours / 2 GB
`qwen3-asr-flash`	Offline	HTTP (OpenAI-compatible)	Unsupported	Supported	Unsupported	Multilingual with dialects	5 minutes / 10 MB
`qwen3-asr-flash-2026-02-10`	Offline	HTTP (OpenAI-compatible)	Unsupported	Supported	Unsupported	Multilingual with dialects	5 minutes / 10 MB
`qwen3-asr-flash-2025-09-08`	Offline	HTTP (OpenAI-compatible)	Unsupported	Supported	Unsupported	Multilingual with dialects	5 minutes / 10 MB

Supported languages: All Qwen-ASR models (qwen3-asr-flash-realtime, qwen3-asr-flash-filetrans, qwen3-asr-flash, and their snapshot versions) share the same language list: Chinese (Mandarin, Sichuanese, Hokkien, Wu, Cantonese), English, Japanese, German, Korean, Russian, French, Portuguese, Arabic, Italian, Spanish, Hindi, Indonesian, Thai, Turkish, Ukrainian, Vietnamese, Czech, Danish, Filipino, Finnish, Icelandic, Malay, Norwegian, Polish, Swedish.

Qwen3.5-Omni / Qwen3-Omni

Model ID	Mode	API	Accuracy boost	Emotion recognition	Speaker diarization	Languages	Max audio duration/size
`qwen3.5-omni-plus-realtime`	Real-time	WebSocket	Prompt context	Supported	Unsupported	Multilingual	2 hours
`qwen3.5-omni-plus`	Offline	HTTP (OpenAI-compatible)	Prompt context	Supported	Unsupported	Multilingual	3 hours / 2 GB
`qwen3.5-omni-flash-realtime`	Real-time	WebSocket	Prompt context	Supported	Unsupported	Multilingual	2 hours
`qwen3.5-omni-flash`	Offline	HTTP (OpenAI-compatible)	Prompt context	Supported	Unsupported	Multilingual	3 hours / 2 GB
`qwen3-omni-flash-realtime`	Real-time	WebSocket	Prompt context	Supported	Unsupported	Multilingual with dialects	2 hours
`qwen3-omni-flash`	Offline	HTTP (OpenAI-compatible)	Prompt context	Supported	Unsupported	Multilingual with dialects	20 minutes / 100 MB

Supported languages: Qwen3.5-Omni / Qwen3-Omni aren't dedicated ASR models. Refer to the user guide and API documentation of each model for supported languages.

Paraformer

Paraformer is an older-generation ASR model family with both real-time and offline variants. Migrate to Fun-ASR or Qwen-ASR when your workload allows.

Model ID	API	Description
`paraformer-realtime-v2`	WebSocket	Real-time recognition; Chinese, English, Japanese, Korean, German, French, Russian
`paraformer-realtime-v1`	WebSocket	Real-time recognition; Chinese, English, Japanese, Korean, German, French, Russian
`paraformer-realtime-8k-v2`	WebSocket	Real-time recognition for 8 kHz telephony; Chinese
`paraformer-realtime-8k-v1`	WebSocket	Real-time recognition for 8 kHz telephony; Chinese
`paraformer-v2`	HTTP	File transcription with speaker diarization; Chinese, English, Japanese, Korean, German, French, Russian
`paraformer-8k-v2`	HTTP	File transcription for 8 kHz telephony; Chinese
`paraformer-v1`	HTTP	File transcription with speaker diarization; Chinese, English, Japanese, Korean, German, French, Russian
`paraformer-8k-v1`	HTTP	File transcription for 8 kHz telephony; Chinese
`paraformer-mtl-v1`	HTTP	File transcription with speaker diarization; multilingual

Supported languages (by version):

paraformer-realtime-v2, paraformer-v2: Chinese (Mandarin, Cantonese, Wu, Hokkien, Northeastern, Gansu, Guizhou, Henan, Hubei, Hunan, Ningxia, Shanxi, Shaanxi, Shandong, Sichuan, Tianjin, Jiangxi, Yunnan, Shanghai), English, Japanese, Korean, German, French, Russian
paraformer-realtime-v1, paraformer-realtime-8k-v2, paraformer-realtime-8k-v1, paraformer-8k-v2, paraformer-8k-v1: Chinese (Mandarin)
paraformer-v1: Chinese (Mandarin), English
paraformer-mtl-v1: Chinese (Mandarin, Cantonese, Wu, Hokkien, Northeastern, Gansu, Guizhou, Henan, Hubei, Hunan, Ningxia, Shanxi, Shaanxi, Shandong, Sichuan, Tianjin), English, Japanese, Korean, Spanish, Indonesian, French, German, Italian, Malay

Others (deprecating)

The following models are scheduled for deprecation and listed for reference only. Migrate to the recommended models as soon as possible.

Model ID	API	Description
`gummy-realtime-v1`	WebSocket	Real-time recognition; Chinese, English, and dialects
`gummy-chat-v1`	WebSocket	Short-audio real-time recognition (1-minute limit); multilingual
`sensevoice-v1`	HTTP	File transcription; multilingual

Audio specifications

The following tables list audio specifications—input method, format, sample rate, and size/duration limits—for real-time and offline modes. For supported languages (including dialects), see the corresponding model family subsection under "All models". Qwen3.5-Omni / Qwen3-Omni models aren't dedicated ASR models. For audio specifications, refer to the user guide and API documentation for each model.

Real-time

Model ID	Input method	Audio format	Sample rate	Size/duration
Fun-ASR-Realtime / Fun-ASR-MTL-Realtime (`fun-asr-realtime`, `fun-asr-mtl-realtime` series)	Binary stream	`pcm`, `wav`, `mp3`, `opus`, `speex`, `aac`, `amr`	Any	Unlimited
Fun-ASR-Flash-8K-Realtime (`fun-asr-flash-8k-realtime` series)	Binary stream	Same as Fun-ASR-Realtime	8 kHz	Unlimited
Qwen-ASR-Realtime (`qwen3-asr-flash-realtime` series)	Binary stream	`pcm`, `opus`	8 kHz, 16 kHz	Unlimited
Paraformer-Realtime (`paraformer-realtime-v2/v1`, `paraformer-realtime-8k-v2/v1`)	Binary stream	Same as Fun-ASR-Realtime	`paraformer-realtime-v2` any rate; `paraformer-realtime-v1` 16 kHz; `paraformer-realtime-8k-*` 8 kHz	Unlimited

All real-time models accept mono (single-channel) input only.

Offline

Model ID	Input method	Audio format	Sample rate	File size/duration
Fun-ASR (`fun-asr`, `fun-asr-mtl` series)	Publicly accessible file URL; 1 URL per request	`aac`, `amr`, `avi`, `flac`, `flv`, `m4a`, `mkv`, `mov`, `mp3`, `mp4`, `mpeg`, `ogg`, `opus`, `wav`, `webm`, `wma`, `wmv`	Any	≤2 GB; ≤12 hours (≤2 hours recommended when speaker diarization is enabled)
Fun-ASR-Flash (`fun-asr-flash-2026-06-15`)	URL / Base64; 1 file per request	`aac`, `amr`, `avi`, `flac`, `flv`, `m4a`, `mkv`, `mov`, `mp3`, `mp4`, `mpeg`, `ogg`, `opus`, `wav`, `webm`, `wma`, `wmv`	Any	≤2 GB; ≤5 minutes
Fun-ASR (`fun-asr-realtime`、`fun-asr-realtime-2026-02-28`)	URL / Base64; 1 file per request	`aac`, `amr`, `avi`, `flac`, `flv`, `m4a`, `mkv`, `mov`, `mp3`, `mp4`, `mpeg`, `ogg`, `opus`, `wav`, `webm`, `wma`, `wmv`	Any	≤2 GB; ≤5 minutes
Paraformer (`paraformer-v2/v1`, `paraformer-mtl-v1`, `paraformer-8k-v2/v1`)	Same as Fun-ASR	Same as Fun-ASR	paraformer-v2 /v1 any rate; `paraformer-8k-*` 8 kHz only ; `paraformer-mtl-v1` 16 kHz or higher	Same as Fun-ASR
Qwen3-ASR-Flash-Filetrans (`qwen3-asr-flash-filetrans` series)	Publicly accessible file URL; 1 URL per request	`aac`, `amr`, `avi`, `flac`, `flv`, `m4a`, `mkv`, `mov`, `mp3`, `mp4`, `mpeg`, `ogg`, `opus`, `wav`, `webm`, `wma`, `wmv`	`pcm` requires 16 kHz; other formats accept any rate (the server resamples to 16 kHz before recognition)	≤2 GB; ≤12 hours
Qwen3-ASR-Flash (`qwen3-asr-flash` series)	URL / Base64 / local file absolute path; 1 file per request	`aac`, `amr`, `avi`, `aiff`, `flac`, `flv`, `mkv`, `mp3`, `mpeg`, `ogg`, `opus`, `wav`, `webm`, `wma`, `wmv`	`pcm` requires 16 kHz; other formats accept any rate (the server resamples to 16 kHz before recognition)	≤10 MB; ≤5 minutes