DGX Spark 使用 Breeze-ASR-25

參考 https://huggingface.co/MediaTek-Research/Breeze-ASR-25

$ uv init breeze-asr

$ cd breeze-asr/

$ rm .python-version

# 在 pyproject.toml 增加下列文件，才可順利安裝 torch 等套件

$ vi pyproject.toml

[[tool.uv.index]]

name = "pytorch-cu130"

url = "https://download.pytorch.org/whl/cu130"

explicit = true  # 關鍵：這會阻止一般套件跑去 PyTorch 倉庫找

[tool.uv.sources]

torch = { index = "pytorch-cu130" }

torchaudio = { index = "pytorch-cu130" }

torchvision = { index = "pytorch-cu130" }

torchcodec = { index = "pytorch-cu130" }

$ uv venv --python 3.13

$ source .venv/bin/activate

$ uv pip install -e .

# 安裝時有順序區別，torch 要比 datasets[audio] 先安裝

$ uv add torch torchaudio torchcodec

$ uv add transformers

$ uv add datasets[audio]

$ uv add accelerate

$ sudo apt update

$ sudo apt install -y ffmpeg libavutil-dev

# 照著文件執行會出現下列錯誤，google AI 和 chatgpt 都建議直接使用 model

KeyError: 'num_frames'

# 若使用下列程式碼，輸出的文件不會分段，全部混再一起

result = processor.batch_decode(

    generated_ids,

    skip_special_tokens=True

)[0]

# 下面為最後的程式，可產生如下的輸出

===== TEXT + TIME =====

0.00s 所說相生這種東西人靠一張嘴

3.28s 通過語言的結構把看官聽眾吸引到故事裡面

7.72s 在演出的時候

9.04s 要求你身上的每個動作都必須要有含義

import torch

import torchaudio

from transformers import WhisperProcessor, WhisperForConditionalGeneration

audio_path = "/home/spark/DiskD/audio_llm/breeze-asr/output.wav"

# load audio

waveform, sr = torchaudio.load(audio_path)

waveform = waveform.mean(dim=0)

if sr != 16000:

    waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)

# load model

processor = WhisperProcessor.from_pretrained("/mnt/models/Breeze-ASR-25")

model = WhisperForConditionalGeneration.from_pretrained(

    "/mnt/models/Breeze-ASR-25"

).to("cuda")

model.eval()

# preprocess

inputs = processor(

    waveform,

    sampling_rate=16000,

    return_tensors="pt"

)

# inference

with torch.no_grad():

    outputs = model.generate(

        inputs.input_features.to("cuda"),

        return_timestamps=True,

        return_dict_in_generate=True,

        output_scores=True,

    )

# --------------------------------------------------

# 5. Decode tokens

# --------------------------------------------------

token_ids = outputs["sequences"][0].tolist()

tokens = processor.tokenizer.convert_ids_to_tokens(token_ids)

# Whisper timestamp token設定

timestamp_begin = processor.tokenizer.convert_tokens_to_ids("<|0.00|>")

time_precision = 0.02  # Whisper: 20ms

current_time = None

buffer = []

print("===== TEXT + TIME =====")

for tid in token_ids:

    # timestamp token

    if tid >= timestamp_begin:

        # 先輸出上一段

        if buffer and current_time is not None:

            text = processor.tokenizer.decode(

                buffer,

                skip_special_tokens=True,

                clean_up_tokenization_spaces=True

            )

            if text.strip():

                print(f"{current_time:.2f}s\t{text}")

        # 更新時間

        current_time = (tid - timestamp_begin) * time_precision

        buffer = []

    else:

        buffer.append(tid)

# flush 最後一段

if buffer and current_time is not None:

    text = processor.tokenizer.decode(

        buffer,

        skip_special_tokens=True,

        clean_up_tokenization_spaces=True

    )

    if text.strip():

        print(f"{current_time:.2f}s\t{text}")