whisper-large-v3-turbo: OpenAI's pruned Whisper large-v3 as a metered transcription API

whisper-large-v3-turbo is OpenAI's speed-optimized cut of Whisper large-v3. On this platform it takes the batch transcription route: one multipart POST with an audio file. The answer is clean text with the duration it bills on. This page gives the model's identity facts with their owners, and the pruning story that makes it fast. It also gives the exact request contract the route enforces, and an honest account of what the turbo pruning costs in accuracy. The route is back in the catalog as of 2026-09-25, served from the platform's own hardware. The restored serving basis prices the lane at the low end of the verified hosted market for this checkpoint, and the rate renders on the pricing page from the same catalog the meter reads. The request shape and the audio-second metering are the same ones the route always ran. The speech-to-text category page carries the current state.

Last verified: 2026-09-25

What it is

whisper-large-v3-turbo is OpenAI's speed-optimized cut of Whisper large-v3. That is the speech-to-text model from the Whisper family. It was released in October 2024. The model card states the relationship in one line. It is a finetuned version of a pruned Whisper large-v3. The exact same model, except the number of decoding layers has been reduced from 32 to 4.

That single sentence carries the whole engineering story. Speech recognition in the Whisper design is an encoder-decoder pass. The encoder turns 30-second audio windows into representations. The decoder turns those into text, token by token. Pruning the decoder from 32 layers to 4 removes most of the per-token work. The audio-understanding side is left untouched. A finetune pass recovers what the cut costs. The result is an 809-million-parameter model, down from large-v3's 1.55 billion. The card describes it as way faster at the expense of a minor quality loss. The detailed comparison sits in OpenAI's own discussion of the release. The approach follows the Distil-Whisper line of work. There a smaller decoder was observed to greatly improve speed with minimal accuracy loss.

The family landscape, because the names blur in search results:

Two naming traps worth clearing. Turbo is not Distil-Whisper. Different authors, same idea, and the turbo card credits the inspiration. And bare "Whisper" in a search result can mean anything from the 2022 base models to the API product. The checkpoint this page documents is exactly openai/whisper-large-v3-turbo.

The signal that this is the workload default rather than a niche cut sits in the card's own numbers. It counts 6.5 million-plus downloads and 3,388 likes at this writing. The weights sit in safetensors under the MIT license. Anyone can download them, and anyone can serve them. That includes this platform.

Use cases

• Batch transcription backlogs. Finished recordings in, text out: voicemail dumps, meeting archives, podcast ingestion, interview corpora. This is the shape turbo was pruned for, where the decode speedup multiplies across thousands of files and the minor accuracy trade is acceptable per-file risk.
• Multilingual ingestion. The card's language list runs to 99 codes with auto-detection from the audio itself, so a mixed-language backlog needs no per-file language tagging. Pin with the language field where you know it; auto-detect where you don't.
• Voice agents and apps that need the transcript, not the stream. Any pipeline that already has the audio as a file (uploads, recordings, fetched media) gets text back in one request with the OpenAI-compatible shape most client libraries already speak.
• Cost-sensitive volume. The 4-layer decoder is the reason turbo-class serving prices well below full large-v3 across the hosting market; if you are evaluating hosts on price per audio minute, turbo is almost always the right tier to compare.

When to route elsewhere on this platform: live streaming while speech happens is a different route. It is documented on the streaming model's page with its own honesty about speaker attribution. Japanese-first workloads have a faster and stronger specialist in hayamimi. The project measured itself 11 to 19 times faster than turbo through faster-whisper on CPU. Its Japanese error rate ran roughly half of turbo's. That comparison is the project's own, dated 2026-08-27, and the hayamimi page carries it.

API usage

Endpoint: POST /v1/audio/transcriptions on https://api.ironstratum.com/v1, authenticated with a bearer API key from the console. The model id is whisper. The request is multipart form data in the OpenAI transcription shape. Client libraries written for that endpoint work with a base URL change. The route serves today, and the contract below is the shape it enforces.

The accepted fields:

curl

curl https://api.ironstratum.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $KEY" \
  -F model=whisper \
  -F file=@meeting.mp3

The response is JSON in the transcription shape. It carries the transcript text, the duration the billing reads, and the fields the selected response format carries.

{
  "text": "Let's start with the quarterly numbers.",
  "duration": 63.4
}

With the language pinned, for when you already know and want to skip detection:

curl https://api.ironstratum.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $KEY" \
  -F model=whisper \
  -F language=en \
  -F file=@meeting.mp3

python

import os
import requests

resp = requests.post(
    "https://api.ironstratum.com/v1/audio/transcriptions",
    headers={"Authorization": f"Bearer {os.environ['KEY']}"},
    files={"file": open("meeting.mp3", "rb")},
    data={"model": "whisper"},
)
resp.raise_for_status()
print(resp.json()["text"])

Limits and errors

The request ceiling is 25 MiB of multipart body. Larger audio follows the client-side chunking discipline. It is covered with the long-recording question below and in depth in the long-audio guide. A request with a missing model bills nothing and returns model_not_found. An unknown form field is rejected at the door. An over-cap body is refused at the door too. Nothing that never reached a model appears on a bill, and the wallet balance is the stop on everything that does.

Benchmarks

What the model card claims, and what it deliberately does not. The card describes turbo's quality cost as a minor loss. It points to the GitHub discussion for the layer-by-layer comparison rather than printing one headline WER delta. That restraint is informative. The honest number depends on the language and the domain, and the card says exactly that in its limitations section. There it notes that the models perform unevenly, with lower accuracy on low-resource languages.

Three measured statements the card does make, each useful in practice:

1. Sequential beats chunked on long audio by up to 0.5 percent WER. Whisper reads 30-second windows, so long audio is either internally chunked or externally split; the internal sequential path is the more accurate one, which argues for the largest single request that fits over aggressive pre-splitting when fidelity matters.
2. Batching parallelizes long files. Multiple audio files transcribe in parallel with a batch size parameter in the Transformers pipeline, the lever for throughput on backlogs when serving locally.
3. The failure modes are documented, not hidden. The sequence-to-sequence architecture is prone to repetitive text, mitigated to a degree by beam search and temperature scheduling but not perfectly, and hallucination risk is worse on low-resource languages. The paper behind the family carries the full evaluation. In hosted practice the mitigations are the prompt field (priming), the language field (pinning, which removes detection errors), and human spot-checks on the languages that matter to you.

For sizing the local option: 809 million parameters is about 1.6 GB in fp16. That follows from the parameter count. That fits modest GPUs comfortably. And faster-whisper's own batched-inference example runs the turbo checkpoint.

Against this platform's other STT routes: the hayamimi page carries that project's own dated comparison. It ran 11-19 times faster than turbo on CPU. Japanese and Cantonese error rates came in roughly half of turbo's, with English and Mandarin slightly behind. The streaming page carries the latency and speaker-attribution trade for live work. This route is the batch generalist: 99 languages, the OpenAI request shape, and the widest tooling of the three.

Getting started

1. Create an API key on the console.
2. Top up the prepaid wallet; spending stops at the balance, never past it.
3. Send the curl request above with a real audio file.

The transcript comes back with the duration it bills on. Every response's usage is visible per call. The rate sits on the pricing page, which renders the same catalog the meter reads. So the number you plan with will be the number the ledger uses.