openai

/

whisper-large-v3

like 3.69k

Follow

OpenAI 2,508

I tried whisper-large-v3 in INT8 and surprisingly the output was better. It transcribed things that FP16 and FP32 missed.
Try it out yourself:

pip install --upgrade transformers datasets[audio] accelerate bitsandbytes torch flash-attn soundfile
huggingface-cli login
mkdir whisper
huggingface-cli download openai/whisper-large-v3 --local-dir ~/whisper --local-dir-use-symlinks False

import torch
import time
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import os
import gc

device = "cuda:0" if torch.cuda.is_available() else "cpu"
model_dir = os.path.expanduser("~/whisper")
audio_file = "audio.mp3"

# Function to transcribe and time
def transcribe(model, processor, dtype, device, audio_file, use_device=True):
    if use_device:
        pipe = pipeline(
            "automatic-speech-recognition",
            model=model,
            tokenizer=processor.tokenizer,
            feature_extractor=processor.feature_extractor,
            torch_dtype=dtype,
            device=device,
        )
    else:
        pipe = pipeline(
            "automatic-speech-recognition",
            model=model,
            tokenizer=processor.tokenizer,
            feature_extractor=processor.feature_extractor,
            torch_dtype=dtype,
        )

    start_time = time.time()
    result = pipe(audio_file)
    end_time = time.time()
    return result["text"], end_time - start_time

# Load and transcribe with fp16
print("Loading model in fp16 precision...")
model_fp16 = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_dir, torch_dtype=torch.float16, low_cpu_mem_usage=True, use_safetensors=True
)
model_fp16.to(device)
processor = AutoProcessor.from_pretrained(model_dir)

print("Transcribing in fp16 precision...")
text_fp16, time_fp16 = transcribe(model_fp16, processor, torch.float16, device, audio_file)
print(f"Transcription (fp16): {text_fp16}")
print(f"Time taken (fp16): {time_fp16:.2f} seconds")

# Unload the fp16 model
del model_fp16
gc.collect()
torch.cuda.empty_cache()

# Load and transcribe with int8
print("Loading model in int8 precision...")
model_int8 = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_dir, 
    load_in_8bit=True, 
    low_cpu_mem_usage=True, 
    use_safetensors=True
)

print("Transcribing in int8 precision...")
text_int8, time_int8 = transcribe(model_int8, processor, torch.float16, device, audio_file, use_device=False)
print(f"Transcription (int8): {text_int8}")
print(f"Time taken (int8): {time_int8:.2f} seconds")

# Compare results
print(f"fp16 vs int8 time difference: {time_fp16 - time_int8:.2f} seconds")