英文

SEW-tiny

SEW by ASAPP Research

基于16kHz采样的语音音频进行预训练的基本模型。使用该模型时,请确保语音输入也以16kHz进行采样。请注意,该模型应在下游任务(如自动语音识别、说话人识别、意图分类、情感识别等)上进行微调。

Paper: Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition

作者:Felix Wu,Kwangyoun Kim,Jing Pan,Kyu Han,Kilian Q. Weinberger,Yoav Artzi

摘要:本文研究了预训练模型在自动语音识别(ASR)中的性能和效率权衡。我们重点研究了wav2vec 2.0,并形式化了几种影响模型性能和效率的架构设计。综合我们的观察,我们引入了SEW(压缩和高效的Wav2vec),这是一种预训练模型架构,在性能和效率两个维度上都有显著改进,适用于各种训练设置。例如,在LibriSpeech的100h-960h半监督设置下,SEW相比wav2vec 2.0实现了1.9倍的推理加速,并且相对于字错误率减少了13.5%。在类似的推理时间下,SEW在不同模型尺寸上将字错误率降低了25-50%。

原始模型可在 https://github.com/asappresearch/sew#model-checkpoints 找到。

用法

要转录音频文件,可以将该模型作为独立的声学模型使用,如下所示:

from transformers import Wav2Vec2Processor, SEWForCTC
from datasets import load_dataset
import soundfile as sf
import torch
 
# load the model and preprocessor
processor = Wav2Vec2Processor.from_pretrained("asapp/sew-tiny-100k-ft-ls100h")
model = SEWForCTC.from_pretrained("asapp/sew-tiny-100k-ft-ls100h")

# load the dummy dataset with speech samples
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
 
# preprocess
input_values = processor(ds[0]["audio"]["array"], return_tensors="pt").input_values  # Batch size 1

# retrieve logits
logits = model(input_values).logits
 
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)

评估

以下代码段展示了如何在LibriSpeech的"clean"和"other"测试数据上评估asapp/sew-tiny-100k-ft-ls100h。

from datasets import load_dataset
from transformers import SEWForCTC, Wav2Vec2Processor
import torch
from jiwer import wer

librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")

model = SEWForCTC.from_pretrained("asapp/sew-tiny-100k-ft-ls100h").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("asapp/sew-tiny-100k-ft-ls100h")

def map_to_pred(batch):
    input_values = processor(batch["audio"][0]["array"], sampling_rate=16000, 
                             return_tensors="pt", padding="longest").input_values
    with torch.no_grad():
        logits = model(input_values.to("cuda")).logits

    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = processor.batch_decode(predicted_ids)
    batch["transcription"] = transcription
    return batch

result = librispeech_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["audio"])

print("WER:", wer(result["text"], result["transcription"]))

结果(WER):

"clean" "other"
10.61 23.74