language: zh datasets:

• superb tags:
• 语音
• 音频
• hubert
• 音频分类 license: apache-2.0 widget:
• example_title: 语音指令 "down" src: https://cdn-media.huggingface.co/speech_samples/keyword_spotting_down.wav
• example_title: 语音指令 "go" src: https://cdn-media.huggingface.co/speech_samples/keyword_spotting_go.wav

Hubert-Large 关键词检测模型

模型描述

此模型为S3PRL的Hubert在SUPERB关键词检测任务上的移植版本。

基础模型采用hubert-large-ll60k，该模型基于16kHz采样的语音音频进行预训练。使用时请确保输入语音同样以16kHz采样。

更多信息请参考SUPERB：语音处理通用性能基准

任务与数据集说明

关键词检测（KS）通过将语音片段分类至预定义的词汇集合，实现预设关键词的识别。该任务通常在设备端实时执行，因此准确性、模型体积和推理速度均为关键指标。SUPERB采用广泛使用的语音指令数据集v1.0，包含10个关键词类别、1个静音类别及1个未知类别（用于误报检测）。

原始模型训练与评估方法详见S3PRL下游任务说明。

使用示例

可通过音频分类流水线调用：

from datasets import load_dataset
from transformers import pipeline

dataset = load_dataset("anton-l/superb_demo", "ks", split="test")

classifier = pipeline("audio-classification", model="superb/hubert-large-superb-ks")
labels = classifier(dataset[0]["file"], top_k=5)

或直接调用模型：

import torch
from datasets import load_dataset
from transformers import HubertForSequenceClassification, Wav2Vec2FeatureExtractor
from torchaudio.sox_effects import apply_effects_file

effects = [["channels", "1"], ["rate", "16000"], ["gain", "-3.0"]]
def map_to_array(example):
    speech, _ = apply_effects_file(example["file"], effects)
    example["speech"] = speech.squeeze(0).numpy()
    return example

# 加载示例数据集并读取音频文件
dataset = load_dataset("anton-l/superb_demo", "ks", split="test")
dataset = dataset.map(map_to_array)

model = HubertForSequenceClassification.from_pretrained("superb/hubert-large-superb-ks")
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("superb/hubert-large-superb-ks")

# 计算注意力掩码并按需标准化波形
inputs = feature_extractor(dataset[:4]["speech"], sampling_rate=16000, padding=True, return_tensors="pt")

logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)
labels = [model.config.id2label[_id] for _id in predicted_ids.tolist()]

评估结果

评估指标为准确率。

BibTeX引用信息

@article{yang2021superb,
  title={SUPERB: Speech processing Universal PERformance Benchmark},
  author={Yang, Shu-wen and Chi, Po-Han and Chuang, Yung-Sung and Lai, Cheng-I Jeff and Lakhotia, Kushal and Lin, Yist Y and Liu, Andy T and Shi, Jiatong and Chang, Xuankai and Lin, Guan-Ting and others},
  journal={arXiv preprint arXiv:2105.01051},
  year={2021}
}