---
title: chinese-wav2vec2-base
canonical_url: "https://www.modelscope.cn/models/TencentGameMate/chinese-wav2vec2-base"
md_url: "https://www.modelscope.cn/models/TencentGameMate/chinese-wav2vec2-base.md"
repository: TencentGameMate/chinese-wav2vec2-base
last_updated: 2025-08-04
license: mit
model_type:
  - wav2vec2
architectures:
  - Wav2Vec2ForPreTraining
parameters: 95.0M
tensor_type:
  - F32
library_name:
  - pytorch
  - transformer
  - safetensors
downloads: 14321
stars: 13
---

# chinese-wav2vec2-base

> chinese-wav2vec2-base - TencentGameMate 在 ModelScope 开源的模型。Pretrained on 10k hours WenetSpeech L subset. More details in TencentGameMate/chinesespeechpretrain

TencentGameMate/chinese-wav2vec2-base 是 ModelScope 魔搭社区上的 95.0M 参数机器学习模型，采用 mit 许可。

- **Repository**: TencentGameMate/chinese-wav2vec2-base
- **License**: mit
- **Parameters**: 95.0M
- **Downloads**: 14321
- **Stars**: 13
- **Last updated**: 2025-08-04

Source: https://www.modelscope.cn/models/TencentGameMate/chinese-wav2vec2-base

---

Pretrained on 10k hours WenetSpeech L subset. More details in  [TencentGameMate/chinese_speech_pretrain](https://github.com/TencentGameMate/chinese_speech_pretrain)

This model does not have a tokenizer as it was pretrained on audio alone. 
In order to use this model speech recognition, a tokenizer should be created and the model should be fine-tuned on labeled text data.

python package:
transformers==4.16.2

```python


import torch
import torch.nn.functional as F
import soundfile as sf
from fairseq import checkpoint_utils

from transformers import (
    Wav2Vec2FeatureExtractor,
    Wav2Vec2ForPreTraining,
    Wav2Vec2Model,
)
from transformers.models.wav2vec2.modeling_wav2vec2 import _compute_mask_indices

model_path=""
wav_path=""
mask_prob=0.0
mask_length=10

feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_path)
model = Wav2Vec2Model.from_pretrained(model_path)

# for pretrain: Wav2Vec2ForPreTraining
# model = Wav2Vec2ForPreTraining.from_pretrained(model_path)

model = model.to(device)
model = model.half()
model.eval()

wav, sr = sf.read(wav_path)
input_values = feature_extractor(wav, return_tensors="pt").input_values
input_values = input_values.half()
input_values = input_values.to(device)

# for Wav2Vec2ForPreTraining
# batch_size, raw_sequence_length = input_values.shape
# sequence_length = model._get_feat_extract_output_lengths(raw_sequence_length)
# mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob=0.0, mask_length=2)
# mask_time_indices = torch.tensor(mask_time_indices, device=input_values.device, dtype=torch.long)

with torch.no_grad():
    outputs = model(input_values)
    last_hidden_state = outputs.last_hidden_state

    # for Wav2Vec2ForPreTraining
    # outputs = model(input_values, mask_time_indices=mask_time_indices, output_hidden_states=True)
    # last_hidden_state = outputs.hidden_states[-1]

```