Short answer: StepAudio 3 ASR Max is StepFun's largest speech recognition model, available on EmpirioLabs through the standard transcription endpoint. Post an audio file to POST /v1/audio/transcriptions with model: "stepaudio-3-asr-max" and you get the transcript back. It detects the spoken language on its own, so there is no language to choose.
Your first request
curl https://api.empiriolabs.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-F "model=stepaudio-3-asr-max" \
-F "file=@meeting.mp3"You can also send JSON with a file_url pointing at a publicly reachable file, or base64 audio in audio_base64.
What it is built for
This is the accuracy tier of the StepFun recognition family. It is built on a large language model, so it uses context rather than sound alone to resolve the parts that ordinary transcription gets wrong:
- Person and place names, drug names, technical terms, and homophones
- Specialist vocabulary across domains such as legal, finance, medical, automotive, and software
- Accented speech, whispering, very fast speech, and mixed-language sentences
- Singing and speech over background music
If you are transcribing ordinary clean speech at volume, StepAudio 2.5 ASR is the cheaper option in the same family and shares this endpoint.
Input formats
Send wav, mp3, ogg, m4a, or pcm. Raw PCM needs codec, rate, bits, and channel alongside it; the container formats carry that information themselves.
Number formatting
Inverse text normalization is on by default, so spoken numbers, dates, and currency come back written out the way a person would type them. Turn it off with enable_itn: false to get the literal words instead.
enable_itn | Transcript |
|---|---|
true (default) | Revenue increased by 12% across all regions last year. |
false | Revenue increased by twelve percent across all regions last year. |
Three things that catch people out
- There is no language setting. ASR Max identifies the language itself and transcribes in it. Japanese audio comes back in Japanese whatever you pass, so do not build a language selector on top of this model. Chinese, English, Japanese, Korean, French, and Spanish are recognized, with everything outside Chinese and English in preview.
- Hotwords and word timestamps are not available here. Both exist on StepAudio 2.5 ASR. If you need a custom vocabulary list or per-word timing, use that model instead.
- Long files are accepted, and billed by their real length. Charging follows the measured duration of the audio you send, with a one second minimum, so trimming silence off the front and back of a recording is worth doing.
Pricing
Billing is per hour of audio, pay as you go, with no subscription and no minimum commitment. The live rate is on the model page and the pricing page.
Try it without writing code
Drop a file into StepAudio 3 ASR Max in the playground and read the transcript straight back. Full parameter reference: docs.empiriolabs.ai/models/stepaudio-3-asr-max.
