← All models

Whisper Large v3 Turbo

Audio Available Full comparison ↗

openai/whisper-large-v3-turbo · by OpenAI · encoder-decoder-transformer

Pricing — 1 offering(s)

Groq

About this offering: OpenAI's open-weights Whisper Large v3 Turbo (a pruned, fine-tuned Large v3) hosted by Groq: $0.04 per hour of audio, the cheapest speech-to-text option Modelglass tracks. Groq lists it at 216× real time with a 12% word error rate, for transcription in 99+ languages; it doesn't do translation. Groq recommends Whisper Large v3 instead for error-sensitive work (Groq speech-to-text docs, checked 2026-10-28).

Audio minutes

  • $0.000667 / minute Current 2026-10-28 → present
    Groq lists Whisper Large v3 Turbo at $0.04 per hour of transcribed audio, about $0.000667 per minute…

    Groq lists Whisper Large v3 Turbo at $0.04 per hour of transcribed audio, about $0.000667 per minute (rounded; $0.04 / 60), with a 10-second minimum billed length per request. The start date is when Modelglass first recorded it; Groq's actual start is earlier.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Whisper Large v3 Turbo fits into a cost-aware routing setup

See how →

About Whisper Large v3 Turbo

Whisper Large v3 Turbo is OpenAI's speed-optimised version of Whisper Large v3, released on 1 October 2024. It keeps Large v3's encoder but has only 4 decoder layers, down from 32. It was fine-tuned for two more epochs on Large v3's multilingual transcription data, without the translation data.

OpenAI reports that it performs similarly to Large v2 across languages, with larger degradation on some languages such as Thai and Cantonese. Groq lists a 12% word error rate for its hosting, against 10.3% for Large v3, at about 216x real time.

This page covers Groq's hosted inference, at $0.04 per hour of audio. Like Large v3 on Groq it is batch / file upload only, with no streaming, diarization or custom vocabulary, and Groq bills a 10-second minimum per request.

Capability profile

How Whisper Large v3 Turbo rates across core capability dimensions, with the task-level evidence behind each rating.

transcription accuracy Moderate

OpenAI says Turbo "performs similarly to large-v2" across languages, with larger degradation on some (Thai, Cantonese), so it gives up most of Large v3's accuracy gain. Groq lists a 12% word error rate against 10.3% for Large v3 and recommends Large v3 for error-sensitive work.

language support Strong

The same 99+ languages as Large v3 for transcription (Groq's model card and speech-to-text docs). It doesn't translate.

speaker diarisation Weak

No native diarization, as with every Whisper checkpoint; it needs external post-processing.

real time streaming Moderate

No native streaming, but Groq runs it at about 216x real time (Groq's speech-to-text docs, checked 2026-10-28), faster than Large v3's 189x, so batch turnaround suits many near-real-time uses.

custom vocabulary Weak

No native custom vocabulary or keyword boosting.

Benchmarks

Benchmark Score Config Source
Word error rate (Groq hosting) 12 % Groq's speech-to-text docs (checked 2026-10-28), which list 10.3% for Whisper Large v3 in the same table. Groq doesn't name the evaluation set. source ↗
Throughput (Groq hosting) 216 x real-time Groq's stated real-time speed factor (checked 2026-10-28); 189x for Whisper Large v3 in the same table. source ↗

Operator guidance

Choose Whisper Large v3 Turbo on Groq for the lowest-cost transcription when roughly large-v2-level accuracy is enough: at $0.04 per hour it costs about a third of Whisper Large v3 on Groq and runs faster. Choose Large v3 for error-sensitive work or translation, and Nova-3 or Universal-2 for native streaming or diarization.

Use cases

Limitations

Citations