changelinglab/librispeech-segment

---
license: cc-by-4.0
language:
- en
pretty_name: LibriSpeech Segment
task_categories:
- automatic-speech-recognition
tags:
- speech
- phone-alignment
- segmentation
- english
size_categories:
- 100K<n<1M
---

# LibriSpeech Segment

English read-speech corpus with **phone-level time alignments** (Montreal
Forced Aligner). Suitable for training and evaluating phone recognition and
phonetic segmentation models.

## Sources

- **Audio**: [LibriSpeech](https://www.openslr.org/12/) (OpenSLR 12) by
Vassil Panayotov, Guoguo Chen, Daniel Povey, Sanjeev Khudanpur (2015).
- **Phone alignments**:
[`anyspeech/librispeech_MFA_alignments`](https://huggingface.co/datasets/anyspeech/librispeech_MFA_alignments).

## Splits

| Split | Utterances |
|-------------------|------------|
| train.clean.100 | 28,538 |
| train.clean.360 | 104,008 |
| train.other.500 | 148,645 |
| dev.clean | 2,703 |
| dev.other | 2,864 |
| test.clean | 2,620 |
| test.other | 2,938 |

Split labels follow the LibriSpeech canonical naming.

## Schema

| Column | Type | Description |
|----------------|----------------------|------------------------------------------------------|
| `utt_id` | string | Utterance id, e.g. `7635-105409-0022` |
| `audio` | Audio(16 kHz) | Embedded waveform bytes (decoded on access) |
| `text` | string | Word-level transcript (uppercase) |
| `phones` | sequence[string] | ARPABET phone tokens |
| `phone_starts` | sequence[float64] | Phone start times in seconds |
| `phone_ends` | sequence[float64] | Phone end times in seconds |
| `language` | string | `eng` (ISO 639-3) |
| `speaker_id` | string | LibriSpeech speaker id |
| `duration` | float64 | Utterance duration in seconds |
| `split` | string | LibriSpeech split label |

## Phone inventory

Phones are ARPABET (e.g. `DH`, `EH`, `R`, `AE`, `OW`). Silence and pauses
are marked with `[SIL]` intervals, kept in the alignment so boundary models
can learn from them. `[UNK]` may also appear for OOV cases.

## License

Released under the **CC BY 4.0** license, matching the original LibriSpeech
audio.

## Citation

```bibtex
@inproceedings{panayotov2015librispeech,
title={Librispeech: an asr corpus based on public domain audio books},
author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
booktitle={2015 IEEE international conference on acoustics, speech and signal processing (ICASSP)},
pages={5206--5210},
year={2015},
organization={IEEE}
}
```