import torch
import librosa
from transformers import WhisperProcessor, WhisperTokenizer, WhisperForConditionalGeneration

# Setups
model_name_or_path = "Oblivion208/whisper-tiny-cantonese"
task = "transcribe"
device = "cuda:0" if torch.cuda.is_available() else "cpu"

model = WhisperForConditionalGeneration.from_pretrained(model_name_or_path).to(device)
tokenizer = WhisperTokenizer.from_pretrained(model_name_or_path, task=task)
processor = WhisperProcessor.from_pretrained(model_name_or_path, task=task)
feature_extractor = processor.feature_extractor
model.config.forced_decoder_ids = None
model.config.suppress_tokens = []

filepath = 'test.wav'
audio, sr = librosa.load(filepath, sr=16000, mono=True)
inputs = processor(audio, sample_rate=sr, return_tensors="pt").to(device)

with torch.inference_mode():
    generated_tokens = model.generate(
        input_features=inputs.input_features,
        return_dict_in_generate=True,
        max_new_tokens=255,
    )
    transcription = tokenizer.batch_decode(
        generated_tokens.sequences, skip_special_tokens=True)
    print(transcription)

Approximate Performance Evaluation

The following models are all trained and evaluated on a single RTX 3090 GPU.

Cantonese Test Results Comparison

MDCC

Model name	Parameters	Finetune Steps	Time Spend	Training Loss	Validation Loss	CER %	Finetuned Model
whisper-tiny-cantonese	39 M	3200	4h 34m	0.0485	0.771	11.10	Link
whisper-base-cantonese	74 M	7200	13h 32m	0.0186	0.477	7.66	Link
whisper-small-cantonese	244 M	3600	6h 38m	0.0266	0.137	6.16	Link
whisper-small-lora-cantonese	3.5 M	8000	21h 27m	0.0687	0.382	7.40	Link
whisper-large-v2-lora-cantonese	15 M	10000	33h 40m	0.0046	0.277	3.77	Link

Common Voice Corpus 11.0

Model name	Original CER %	w/o Finetune CER %	Jointly Finetune CER %
whisper-tiny-cantonese	124.03	66.85	35.87
whisper-base-cantonese	78.24	61.42	16.73
whisper-small-cantonese	52.83	31.23	/
whisper-small-lora-cantonese	37.53	19.38	14.73
whisper-large-v2-lora-cantonese	37.53	19.38	9.63