File size: 3,824 Bytes
57a71c9
6486b01
ee9020e
 
 
14de29e
ee9020e
 
 
 
 
57a71c9
ee9020e
 
 
 
 
 
 
 
 
 
 
14de29e
 
ee9020e
57a71c9
ee9020e
 
 
3e34d55
ee9020e
 
3e34d55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14de29e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee9020e
 
 
14de29e
ee9020e
14de29e
ee9020e
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
---
language: zh
datasets:
- aishell1
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large 53 - Chinese (zh-CN), by Yue Qin
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: AISHELL-1 zh-CN
      type: aishell1
      args: zh-CN
    metrics:
       - name: Test WER
         type: wer
         value: 7.04
---

# Wav2Vec2-Large-XLSR-53-Chinese-zh-CN-aishell1

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Chinese using the [AISHELL-1](https://github.com/kaldi-asr/kaldi/tree/master/egs/aishell) dataset.
When using this model, make sure that your speech input is sampled at 16kHz.

## Usage

The model can be used directly (without a language model) as follows:

```python
import torch
import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

device = "cuda:0" if torch.cuda.is_available() else "cpu"

processor = Wav2Vec2Processor.from_pretrained(
    'qinyue/wav2vec2-large-xlsr-53-chinese-zn-cn-aishell1')
model = Wav2Vec2ForCTC.from_pretrained(
    'qinyue/wav2vec2-large-xlsr-53-chinese-zn-cn-aishell1').to(device)

filepath = 'test.wav'
audio, sr = librosa.load(filepath, sr=16000, mono=True)
inputs = processor(audio, sample_rate=16000, return_tensors="pt").to(device)
with torch.no_grad():
    logits = model(inputs.input_values,
                   attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
pred_str = processor.decode(predicted_ids[0])

print(pred_str)
```

## Evaluation

```python
wer_metric = load_metric("wer")

def compute_metrics(pred):
    pred_logits = pred.predictions
    pred_ids = np.argmax(pred_logits, axis=-1)

    pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id

    pred_str = processor.batch_decode(pred_ids, spaces_between_special_tokens=True)
    label_str = processor.batch_decode(pred.label_ids, group_tokens=False, spaces_between_special_tokens=True)

    wer = wer_metric.compute(predictions=pred_str, references=label_str)

    return {"wer": wer}
```

## Results

| Reference  | Prediction |
| ------------- | ------------- |
| 据 伟 业 我 爱 我 家 市 场 研 究 院 测 算 | 据 北 业 我 爱 我 家 市 场 研 究 院 测 算 |
| 七 月 北 京 公 积 金 贷 款 成 交 量 提 升 了 百 分 之 五 | 七 月 北 京 公 积 金 贷 款 成 交 量 提 升 了 百 分 之 五 |
| 培 育 门 类 丰 富 层 次 齐 用 的 综 合 利 用 产 业 | 培 育 门 类 丰 富 层 资 集 业 的 综 合 利 用 产 业 |
| 我 们 迎 来 了 赶 超 发 达 国 家 的 难 得 机 遇 | 我 们 迎 来 了 赶 超 发 达 国 家 的 单 得 机 遇 |
| 坚 持 基 本 草 原 保 护 制 度 | 坚 持 基 本 草 员 保 护 制 度 |
| 强 化 水 生 生 态 修 复 和 建 设 | 强 化 水 生 生 态 修 复 和 建 设 |
| 温 州 两 男 子 为 争 女 人 驾 奔 驰 宝 马 街 头 四 次 对 撞 | 温 州 两 男 子 为 争 女 人 架 奔 驰 宝 马 接 头 四 次 对 重 |
| 她 表 示 应 该 是 吃 吃 饭 看 电 影 之 类 的 | 他 表 示 一 的 是 吃 吃 饭 看 电 影 之 理 |
| 加 强 畜 禽 遗 传 资 源 和 农 业 野 生 植 物 资 源 保 护 | 加 强 续 紧 遗 传 资 源 和 农 业 野 生 职 物 资 源 保 护 |
| 两 人 都 是 依 赖 电 话 沟 通 | 两 人 都 是 依 赖 电 话 沟 通 |

**Test Result**:

In the table below I report the Word Error Rate (WER) of the model on the AISHELL-1 test dataset.

| Model | WER | WER-with-LM |
| ------------- | ------------- | ------------- |
| qinyue/wav2vec2-large-xlsr-53-chinese-zn-cn-aishell1 | **7.04%** | **3.96%** |