LibriSpeech

Overview

LibriSpeech is a large-scale corpus of approximately 1,000 hours of read English speech derived from audiobooks. It is one of the most widely used benchmarks for evaluating automatic speech recognition (ASR) systems.

Task Description

Task Type: Automatic Speech Recognition (ASR)
Input: Audio recordings of read English speech from audiobooks
Output: Transcribed text
Language: English

Key Features

1,000 hours of high-quality read speech
Derived from LibriVox audiobooks (public domain)
Clean and "other" test sets for varying difficulty
Widely used baseline for ASR research
Standardized evaluation protocol

Evaluation Notes

Default configuration uses test_clean split
Primary metric: Word Error Rate (WER)
Text normalization applied during evaluation
Prompt: "Please recognize the speech and only output the recognized content"
Metadata includes audio ID and duration information

Properties

Property	Value
Benchmark Name	`librispeech`
Dataset ID	lmms-lab/Librispeech-concat
Paper	N/A
Tags	`Audio`, `SpeechRecognition`
Metrics	`wer`
Default Shots	0-shot
Evaluation Split	`test_clean`

Data Statistics

Metric	Value
Total Samples	87
Prompt Length (Mean)	67 chars
Prompt Length (Min/Max)	67 / 67 chars

Audio Statistics:

Metric	Value
Total Audio Files	87
Audio per Sample	min: 1, max: 1, mean: 1
Formats	wav

Sample Example

Subset: default

{
  "input": [
    {
      "id": "fd0309e6",
      "content": [
        {
          "text": "Please recognize the speech and only output the recognized content:"
        },
        {
          "audio": "[BASE64_AUDIO: wav, ~22.7MB]",
          "format": "wav"
        }
      ]
    }
  ],
  "target": "Eleven o'clock had struck it was a fine clear night they were the only persons on the road and they sauntered leisurely along to avoid paying the price of fatigue for the recreation provided for the toledans in their valley or on the banks of ... [TRUNCATED] ...  less surprised than they and the better to assure himself of so wonderful a fact he begged leocadia to give him some token which should make perfectly clear to him that which indeed he did not doubt since it was authenticated by his parents.",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "audio_id": "5639-40744",
    "audio_duration": 496.6899719238281
  }
}

Note: Some content was truncated for display.

Prompt Template

Prompt Template:

Please recognize the speech and only output the recognized content:

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets librispeech \
    --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=['librispeech'],
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)

3.3 KiB Raw Blame History