Skip to main content
The StepAudio 3 ASR family provides high-accuracy, low-latency speech transcription. StepAudio 3 ASR Max is StepFun’s largest ASR model to date. Built on a large language model, it fuses high-precision acoustic modeling with that model’s context understanding, knowledge, and reasoning—moving speech recognition from recognizing sound toward understanding context.

​
Model information

​
Core capabilities

  • Context understanding: combines the audio signal, contextual relationships, and domain knowledge to improve recognition of complex content such as person and place names, drug names, technical terms, and homophones.
  • Cross-language, cross-scenario recognition: covers Chinese, English, dialects, mixed Chinese-English, long audio, and vertical domains, improving over the previous generation.
  • Professional-domain recognition: understands specialized expressions using domain knowledge across 20+ vertical scenarios such as sports, pharmaceuticals, chemicals, automotive, legal, finance, and software development.
  • Complex audio handling: supports whispering, very fast speech, slurred liaison, dragged pronunciation, ambient noise, singing, and background music.
  • Song and singing transcription: handles complex singing scenarios such as a cappella and accompanied singing, transcribing song content directly.
  • Low misrecognition: 99% accuracy on no-speech tests, reducing erroneous transcription caused by ambient sound, background music, and silent segments.

​
Benchmark highlights

  • Contextual reasoning: 0.57% error rate—down 64.2% from StepAudio 2.5 ASR’s 1.59%, and down 43.0% from Doubao ASR 2.0’s 1.00%.
  • Vertical-domain proper nouns: covering 20+ domains such as sports, pharmaceuticals, chemicals, automotive, legal, finance, and software development, with a 3.81% average error rate.
  • Across-the-board gains over the previous generation: average error rates down 9.3% for general Chinese and 25.0% for English, 22.3% for dialects, 42.1% for vertical domains, and 27.9% for mixed Chinese-English.
  • Long-context capability: contextual-reasoning error rate down 64.2% versus the previous generation.
  • Versus leading commercial models: compared with Doubao ASR 2.0, average error rates down 18.6% for general Chinese, 60.4% for English, 23.3% for heavy accents, 40.3% for long audio, and 43.0% for contextual reasoning.
  • Real-world complex audio: 3.97% CER on the AISHELL-6 whisper test, better than the previous generation’s 5.36%; stable recognition across very fast speech, slurring, liaison, and dragging tests.

​
Use cases

StepAudio 3 ASR Max is offered through the API and can be used for meeting minutes, video subtitles, live-stream content understanding, smart customer service, in-vehicle interaction, medical records, financial services, and professional content production—making spoken content more accurately understood, searched, and used.

​
API endpoint

Speech Recognition (Streaming Output)

POST /v1/audio/asr/sse
Submit audio in one request; recognition text is returned as an SSE stream.

Audio models overview

Back to the Audio 3 model overview.

Full pricing details

Billing rules for all speech, text, and image models.

Speech Recognition API

Request parameters, response events, and interface details.

Voice Studio

Try the recognition of StepAudio 3 ASR online.