Speech-to-Text API
Whisper Large v3 and NVIDIA Parakeet on the OpenAI /v1/audio/transcriptions endpoint. Upload audio, get text with timestamps. $0.002 per minute, 100 MB files, no daily cap.
Models
| Model | Languages | Price | Measured speed | Status |
|---|---|---|---|---|
| Whisper Large v3 OpenAI | 100 languages | $0.002 / min | 0.047× real time | Available |
| Parakeet TDT 0.6B v3 NVIDIA | 25 languages | $0.002 / min | 0.036× real time | Available |
| Voxtral Small 24B Mistral AI | 8 languages, Auto language | $0.004 / min | 0.09× real time | Available |
| Voxtral Mini 3B Mistral AI | 8 languages, Auto language | $0.0013 / min | 0.058× real time | Available |
Quickstart
curl https://api.inferenceapis.com/v1/audio/transcriptions \
-H "Authorization: Bearer $INFERENCE_API_KEY" \
-F "model=openai/whisper-large-v3" \
-F "file=@meeting.mp3" \
-F "response_format=verbose_json"import os
from openai import OpenAI
client = OpenAI(base_url="https://api.inferenceapis.com/v1", api_key=os.environ["INFERENCE_API_KEY"])
with open("meeting.mp3", "rb") as f:
result = client.audio.transcriptions.create(model="openai/whisper-large-v3", file=f, response_format="verbose_json")
print(result.text)
for s in result.segments:
print(f"{s.start:6.1f}s {s.text}")Parameters, Node.js examples and response shapes are on the Whisper and Parakeet pages.
Pricing versus alternatives
Per minute of audio. External prices are public pay-as-you-go list rates as of September 2026; volume plans differ.
| Provider / model | Price per minute | 10 hours of audio | Notes |
|---|---|---|---|
| Inference APIs — Whisper Large v3 | $0.002 | $1.20 | 100 MB files, word timestamps, speaker diarization at no extra charge, SRT/VTT; accuracy measured per language |
| Inference APIs — Parakeet TDT 0.6B v3 | $0.002 | $1.20 | 25 languages, fastest option here |
| Groq — whisper-large-v3-turbo | $0.0007 | $0.42 | Free tier capped at 28,800 audio-seconds/day; Developer-tier upgrades paused |
| AssemblyAI — Universal | $0.0025 | $1.50 | Diarization and speech understanding included |
| OpenAI — whisper-1 | $0.006 | $3.60 | 25 MB file limit |
| Deepgram — Nova-3 | $0.0077 | $4.62 | Streaming, diarization, keyterms |
Side-by-side pages: Whisper here vs OpenAI whisper-1, vs Groq, Whisper vs Parakeet.
What people build with it
- Meeting and call notes — transcribe recordings, then summarise with a chat model on the same key.
- Subtitles —
response_format=srtstraight into a video pipeline. - Who said what —
diarize=truelabels each speaker turn in interviews, podcasts and calls, in JSON, plain text or subtitle form. How it performed on our test clips. - Voice input — dictation apps, voice commands and assistants that were on a free tier that ran out. Open WebUI can switch its STT engine to this endpoint with three environment variables.
Frequently asked questions
Is it compatible with the OpenAI transcription API?
Yes. POST /v1/audio/transcriptions takes the same multipart fields (file, model, language, prompt, response_format, timestamp_granularities). The OpenAI SDKs work by setting base_url; whisper-1 is accepted as an alias for Whisper Large v3.
What is the file size limit?
100 MB per request (OpenAI's endpoint allows 25 MB). Supported containers: mp3, mp4, m4a, wav, webm, ogg, flac.
Do I get timestamps?
Yes. response_format=verbose_json returns language, duration and, with timestamp_granularities[]=word, a start and end time for every word. srt and vtt return ready-made subtitle files with cues broken at sentence ends and pauses.
How is it billed?
Per minute of audio, measured from the uploaded file with ffprobe before it is transcribed. A 90-second clip is billed as 1.5 minutes.
Which languages are supported?
Whisper Large v3 accepts 100 language codes and Parakeet TDT 0.6B v3 covers 25 European languages. Accepting a language is not the same as transcribing it well, so we measured both models on public test clips: about a third came out good or excellent, and about a third are not usable, including a few that return empty transcripts. See accuracy by language before you commit.
Is speaker diarization available?
Yes, on Whisper Large v3: add diarize=true (and optionally min_speakers / max_speakers). Every word gets a speaker, segments become one entry per speaker turn, and text, srt and vtt output carry SPEAKER_00: labels. Same $0.002 per minute. Measured results and limits are on the diarization page.
Is real-time streaming available?
No. Both models transcribe recorded files. For live audio, Deepgram is in the comparison table above.
Is there a free tier or a daily cap?
There is no free tier and no daily cap on requests or audio minutes: usage is billed per minute from a prepaid balance, and you can add as little as you like.