Disclosure: This article was written with AI assistance and reviewed by EmpirioLabs AI.
Short answer: on the same 12.6-second voicemail, Qwen Audio 3.1 ASR cost the least and returned first, and wrote every number as digits. Qwen Audio 3.0 ASR wrote the numbers out as words, StepAudio 3 ASR Max was close to 3.1 ASR, and Deepgram Nova-3 cost the most and split the invoice number. All four run on EmpirioLabs through the same transcription endpoint. Read each transcript below; we are not declaring a winner.
The voicemail
The words spoken: "Hi, this is Dr. Amara Okafor calling about invoice 4417. The balance of two hundred and eighty-six dollars was due on March 14th. Please call me back at 555 0142 before Friday."
The four transcripts
Qwen Audio 3.0 ASR: "Hi, this is Dr. Amara Okafor calling about Invoice four thousand four hundred seventeen. The balance of two hundred eighty-six dollars was due on March the fourteenth. Please call me back at five hundred fifty-five zero one four two before Friday."
Qwen Audio 3.1 ASR: "Hi, this is Dr. Amara Okafor calling about invoice 4417. The balance of $286 was due on March 14th. Please call me back at 555-0142 before Friday."
StepAudio 3 ASR Max: "Hi, this is Doctor Amara Okafor calling about invoice 4417. The balance of $286 was due on March the 14th. Please call me back at 555 0142 before Friday."
Deepgram Nova-3: "Hi. This is Doctor. Amara Okafer calling about invoice 4400 and 17. The balance of $286 was due on March 14. Please call me back at 555-0142 before Friday."
Results at a glance
| Qwen Audio 3.0 ASR | Qwen Audio 3.1 ASR | StepAudio 3 ASR Max | Deepgram Nova-3 | |
|---|---|---|---|---|
| Maker | Alibaba | Alibaba | StepFun | Deepgram |
| Billing unit | Hours of audio, by the second | Tokens | Hours of audio | Minutes of audio |
| Price for this recording | $0.0009 | $0.0001 | $0.0008 | $0.0029 |
| Time to the transcript | 5.9 seconds | 2.5 seconds | 2.8 seconds | 4.5 seconds |
| Numbers | As words | As digits | As digits | As digits |
| Name | Amara Okafor | Amara Okafor | Amara Okafor | Amara Okafer |
How we ran it
The voicemail was synthesized once with Qwen Audio 3.1 TTS Plus and sent to each model as the same MP3 file through POST /v1/audio/transcriptions, with every setting at its default and no prompt or vocabulary. One shot each, on November 2, 2026. Times run from the upload to the completed job, polled every 1.5 seconds. Prices are the amounts charged for these exact requests.
What to look for
Numbers. The voicemail holds an invoice number, an amount, a date and a phone number. Qwen Audio 3.0 ASR spelled each one out; the other three wrote digits, and Deepgram Nova-3 heard the invoice number as "4400 and 17".
Where 3.0 fits. Qwen Audio 3.0 ASR is priced per hour of audio and covers the major Chinese dialect groups and 30 languages; test it on the speech you need to transcribe.
Run the same test on EmpirioLabs
curl https://api.empiriolabs.ai/v1/audio/transcriptions \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-F model=qwen-audio-3-0-asr \
-F file=@voicemail.mp3The request returns a job; read the transcript from GET /v1/jobs/<job_id> once it completes. Swap model to qwen-audio-3-1-asr, stepaudio-3-asr-max or deepgram-nova-3 to run the others, or upload a recording in the playground.
Frequently asked questions
Were the transcripts edited or retried?
No. Each transcript is the model's first result for this recording, unedited.
Can a prompt fix the names and numbers?
No prompt was used here. Qwen Audio 3.0 ASR Stream and Qwen Audio 3.1 ASR Stream take a prompt of names and terms on a live socket; see How to Use Qwen Audio 3.0 ASR and ASR Stream.
Which one should I use?
Transcribe a recording like yours with each. The same request works for all four with only the model changed.
