google/fleurs
Viewer • Updated • 768k • 206k • 515
This repository contains the CTranslate2 converted version of Dafisns/whisper-turbo-multilingual-fleurs.
It is optimized for lightning-fast inference and low memory usage using the faster-whisper library. The model was fine-tuned on a mix of Indonesian and English datasets, including Google FLEURS, Common Voice 22.0, and EdAcc (for Indonesian-accented English).
The following Word Error Rates (WER) were achieved by the original model on the combined test sets:
| Language | Dataset Composition | WER (%) |
|---|---|---|
| English | Fleurs + Common Voice + EdAcc | 9.09% |
| Indonesian | Fleurs + Common Voice | 6.97% |
You need to install the faster-whisper library to use this model efficiently:
pip install faster-whisper
from faster_whisper import WhisperModel
# Use 'cuda' for GPU or 'cpu' for CPU
# 'float16' is recommended for GPU, 'int8' for CPU
model_id = "Dafisns/whisper-turbo-multilingual-fleurs-ct2"
model = WhisperModel(model_id, device="cuda", compute_type="float16")
# Transcribe audio file
# Setting language='id' ensures the model focuses on Indonesian
segments, info = model.transcribe("audio.mp3", beam_size=1, language="id")
print(f"Detected language '{info.language}' with probability {info.language_probability}")
for segment in segments:
print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")
Base model
openai/whisper-large-v3