CoolFace
Modelpublic

soniqo/Whisper-Large-v3-Turbo-ONNX

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes9downloads
Model Card

Whisper Large v3 Turbo - ONNX

Official OpenAI Whisper large-v3-turbo checkpoint exported as ONNX encoder/decoder graphs for speech-core and sherpa-onnx-compatible runtimes.

Part of the soniqo.audio speech toolkit. This is the ONNX Runtime bundle used by speech-core for server, desktop, and Android-style runtimes; the graphs remain sherpa-onnx compatible. Browse ONNX bundles in the soniqo ONNX collection.

Model

Sourceopenai/whisper-large-v3-turbo
Export formatONNX for sherpa-onnx
VariantsFP32, INT8, FP16
Runtimespeech-core OnnxWhisperStt / ONNX Runtime; sherpa-onnx compatible
Total artifact size5619.69 MiB

The FP16 graphs use external *.onnx.data files. Large-style bundles also use external FP32 *.weights files. Keep external-data files beside their matching .onnx files.

Files

FileSizeDescription
turbo-decoder.fp16.onnx0.59 MBWhisper decoder graph
turbo-decoder.fp16.onnx.data302.88 MBExternal tensor data for the adjacent ONNX graph
turbo-decoder.int8.onnx344.34 MBWhisper decoder graph
turbo-decoder.onnx0.97 MBWhisper decoder graph
turbo-decoder.weights605.76 MBExternal tensor data for the adjacent ONNX graph
turbo-encoder.fp16.onnx0.61 MBWhisper encoder graph
turbo-encoder.fp16.onnx.data1239.93 MBExternal tensor data for the adjacent ONNX graph
turbo-encoder.int8.onnx643.37 MBWhisper encoder graph
turbo-encoder.onnx0.6 MBWhisper encoder graph
turbo-encoder.weights2479.86 MBExternal tensor data for the adjacent ONNX graph
turbo-tokens.txt0.78 MBTokenizer tokens for speech-core and sherpa-onnx-compatible runtimes

Usage

Use with speech-core's native ONNX Whisper runtime:

cpp
#include <speech_core/models/onnx_whisper_stt.h>

speech_core::OnnxWhisperStt stt(
    "turbo-encoder.int8.onnx",
    "turbo-decoder.int8.onnx",
    "turbo-tokens.txt");

auto result = stt.transcribe(audio, length, 16000);

The same encoder/decoder/token files can also be loaded by sherpa-onnx:

python
import sherpa_onnx

recognizer = sherpa_onnx.OfflineRecognizer.from_whisper(
    encoder="turbo-encoder.fp16.onnx",
    decoder="turbo-decoder.fp16.onnx",
    tokens="turbo-tokens.txt",
    language="en",
    task="transcribe",
    provider="cpu",
)

Full FLEURS WER / RTF

M5 Pro, local FLEURS test split, sherpa-onnx CPU provider, language hint supplied, temperature fallback disabled.

DatasetINT8 WER / RTFFP16 WER / RTF
fleurs-en_us5.32% / 0.2395.27% / 0.189
fleurs-fr_fr6.84% / 0.1886.57% / 0.215
fleurs-ar_eg14.53% / 0.20114.30% / 0.215

Note: the French split contains one 33.54 s clip; sherpa-onnx Whisper processes the first 30 s.

Links