英文

S2T-LARGE-LIBRISPEECH-ASR

s2t-large-librispeech-asr是一个用于自动语音识别(ASR)的Speech to Text Transformer(S2T)模型。该S2T模型是在 this paper 中提出并在 this repository 中发布的。

模型描述

S2T是一个端到端序列到序列的转换器模型。它使用标准的自回归交叉熵损失进行训练,并以自回归的方式生成转录。

预期用途和限制

该模型可用于端到端的语音识别(ASR)。请参阅 model hub 以查找其他S2T检查点。

如何使用

由于这是一个标准的序列到序列转换器模型,您可以使用generate方法通过将语音特征传递给模型来生成转录。

注意:Speech2TextProcessor对象使用 torchaudio 来提取滤波器组特征。在运行此示例之前,请确保安装torchaudio软件包。

您可以使用pip install transformers"[speech, sentencepiece]"命令将其安装为额外的语音依赖项,或者使用pip install torchaudio sentencepiece命令分别安装这些软件包。

import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf

model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-large-librispeech-asr")
processor = Speech2Textprocessor.from_pretrained("facebook/s2t-large-librispeech-asr")

def map_to_array(batch):
    speech, _ = sf.read(batch["file"])
    batch["speech"] = speech
    return batch

ds = load_dataset(
    "patrickvonplaten/librispeech_asr_dummy",
    "clean",
    split="validation"
)
ds = ds.map(map_to_array)

input_features = processor(
    ds["speech"][0],
    sampling_rate=16_000,
    return_tensors="pt"
).input_features  # Batch size 1
generated_ids = model.generate(input_ids=input_features)

transcription = processor.batch_decode(generated_ids)
在LibriSpeech测试集上的评估

以下脚本展示了如何在 LibriSpeech 的“clean”和“other”测试数据集上评估此模型。

from datasets import load_dataset, load_metric
from transformers import Speech2TextForConditionalGeneration, Speech2TextProcessor
import soundfile as sf

librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")  # change to "other" for other test dataset
wer = load_metric("wer")

model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-large-librispeech-asr").to("cuda")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-large-librispeech-asr", do_upper_case=True)

def map_to_array(batch):
    speech, _ = sf.read(batch["file"])
    batch["speech"] = speech
    return batch

librispeech_eval = librispeech_eval.map(map_to_array)

def map_to_pred(batch):
    features = processor(batch["speech"], sampling_rate=16000, padding=True, return_tensors="pt")
    input_features = features.input_features.to("cuda")
    attention_mask = features.attention_mask.to("cuda")

    gen_tokens = model.generate(input_ids=input_features, attention_mask=attention_mask)
    batch["transcription"] = processor.batch_decode(gen_tokens, skip_special_tokens=True)
    return batch

result = librispeech_eval.map(map_to_pred, batched=True, batch_size=8, remove_columns=["speech"])

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

结果(WER):

"clean" "other"
3.3 7.5

训练数据

S2T-LARGE-LIBRISPEECH-ASR是在 LibriSpeech ASR Corpus 上训练的,该数据集包含约1000小时的16kHz英语朗读音频。

训练过程

预处理

语音数据通过从WAV / FLAC音频文件中自动提取符合Kaldi规范的80通道对数梅尔滤波器组特征进行预处理,可以使用PyKaldi或torchaudio完成。对每个例子都应用了基于句子的CMVN(倒谱平均值和方差归一化)。

文本使用SentencePiece进行小写和分词处理,词汇表大小为10,000。

训练

模型使用标准的自回归交叉熵损失进行训练,并使用 SpecAugment 进行优化。编码器接收语音特征,解码器自回归生成转录。

BibTeX条目和引用信息

@inproceedings{wang2020fairseqs2t,
  title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
  author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
  booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
  year = {2020},
}