whisper-turbo-az-calls-test

Test model: openai/whisper-large-v3-turbo fine-tuned (LoRA, merged) on Azerbaijani call-center speech (8 kHz telephone audio).

Usage

import librosa
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "vrashad/whisper-turbo-az-calls-test"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id, torch_dtype=dtype).to(device).eval()

audio, _ = librosa.load("audio.wav", sr=16000, mono=True)   # one channel, up to 30 s
features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.inference_mode():
    ids = model.generate(features.to(device, dtype), language="az", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

For longer audio, split it at pauses into pieces of up to 30 s and transcribe them in a batch.