language: zh tags:

• 语音转文本 license: apache-2.0

wav2vec2-base-da-ft-nst

这是由Alvenir基于公开NST数据集微调的丹麦语wav2vec2模型。该模型训练于16kHz采样率，请确保您的数据采样率一致。

本模型使用fairseq框架训练，后转换为huggingface/transformers格式。

Alvenir始终乐于协助您的开源ASR项目，提供定制领域优化服务或高级模型支持。;-)

使用方式

import soundfile as sf
import torch

from transformers import Wav2Vec2CTCTokenizer, Wav2Vec2Tokenizer, Wav2Vec2Processor, \
    Wav2Vec2ForCTC


def get_tokenizer(model_path: str) -> Wav2Vec2CTCTokenizer:
    return Wav2Vec2Tokenizer.from_pretrained(model_path)


def get_processor(model_path: str) -> Wav2Vec2Processor:
    return Wav2Vec2Processor.from_pretrained(model_path)


def load_model(model_path: str) -> Wav2Vec2ForCTC:
    return Wav2Vec2ForCTC.from_pretrained(model_path)


model_id = "Alvenir/wav2vec2-base-da-ft-nst"

model = load_model(model_id)
model.eval()
tokenizer = get_tokenizer(model_id)
processor = get_processor(model_id)

audio_file = "<音频文件路径/audio.wav>"

audio, _ = sf.read(audio_file)

input_values = processor(audio, return_tensors="pt", padding="longest", sampling_rate=16_000).input_values
with torch.no_grad():
    logits = model(input_values).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
print(transcription)

基准测试结果

以下是在丹麦公开数据集上的基准测试表现。