WenetSpeech

Overview

WenetSpeech is a large-scale Mandarin Chinese speech corpus with over 10,000 hours of multi-domain transcribed audio data, designed for speech recognition research.

Task Description

Task Type: Automatic Speech Recognition (ASR)
Input: Audio recordings with Mandarin Chinese speech
Output: Transcribed text in Chinese
Domain: Multi-domain (internet, meeting)

Key Features

Large-scale Mandarin Chinese speech corpus (10,000+ hours)
Multi-domain coverage: internet content, meetings
High-quality transcriptions
Suitable for evaluating Chinese ASR systems
Supports mixed Chinese-English text evaluation

Evaluation Notes

Default configuration uses test_meeting split
Subsets by domain: dev (development), test_meeting (meeting domain)
Primary metric: MER (Mixed Error Rate)
MER tokenizes Chinese characters individually and English words as whole tokens
Prompt: "Please listen to the audio and transcribe what you hear"

Properties

Property	Value
Benchmark Name	`wenet_speech`
Dataset ID	lmms-lab/WenetSpeech
Paper	N/A
Tags	`Audio`, `SpeechRecognition`
Metrics	`mer`
Default Shots	0-shot
Evaluation Split	`test_meeting`

Data Statistics

Metric	Value
Total Samples	22,195
Prompt Length (Mean)	161 chars
Prompt Length (Min/Max)	161 / 161 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
`dev`	13,825	161	161	161
`test_meeting`	8,370	161	161	161

Audio Statistics:

Metric	Value
Total Audio Files	22,195
Audio per Sample	min: 1, max: 1, mean: 1
Formats	wav

Sample Example

Subset: dev

{
  "input": [
    {
      "id": "c30c80b4",
      "content": [
        {
          "text": "Please listen to the audio and transcribe what you hear. Please only provide the transcription without any additional commentary. Do not include any punctuation."
        },
        {
          "audio": "[BASE64_AUDIO: wav, ~175.3KB]",
          "format": "wav"
        }
      ]
    }
  ],
  "target": "对我做了介绍啊那么我想说的是呢大家如果对我的研究感兴趣呢嗯",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "text": "对我做了介绍啊那么我想说的是呢大家如果对我的研究感兴趣呢嗯"
  }
}

Prompt Template

Prompt Template:

Please listen to the audio and transcribe what you hear. Please only provide the transcription without any additional commentary. Do not include any punctuation.

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets wenet_speech \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['wenet_speech'],
    dataset_args={
        'wenet_speech': {
            # subset_list: ['dev', 'test_meeting']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)

3.5 KiB Raw Permalink Blame History