---
language:
  - en
license: cc-by-4.0
tags:
  - ultrasound
  - speech
  - mel-spectrogram
  - tongue-imaging
  - silent-speech
  - tal-corpus
task_categories:
  - audio-to-audio
---

# TaL Corpus — Preprocessed HDF5 Dataset

Preprocessed HDF5 files derived from the
[TaL corpus](https://ultrasuite.github.io/data/tal_corpus/), containing
synchronized ultrasound tongue imaging (UTI) and mel spectrogram frames
for four speakers. Intended for use with
[ult-to-speech-pytorch](https://github.com/ibrahimkhaliloglu/ult-to-speech-pytorch).

## Dataset Description

Each speaker is stored in a single `.h5` file with the following layout:

```
{speaker}.h5
  /ult          (N, 64, 128)   uint8    ultrasound frames, normalized to [-1, 1] at load time
  /mel          (N, 80)        float32  log mel spectrogram, HiFi-GAN compatible
  /boundaries   (R, 2)         int64    [start, end) frame indices per utterance
  /filenames    (R,)           str      utterance base names
```

Please cite the original corpus if you use this dataset:

> Ribeiro, M. S., Sanger, J., Zhang, J.-X., Eshky, A., Wrench, A., Richmond, K., & Renals, S. (2021).
> TaL: a synchronised multi-speaker corpus of ultrasound tongue imaging, audio, and lip videos.
> *Proceedings of the IEEE Workshop on Spoken Language Technology (SLT)*, Shenzhen, China.
> [[paper](https://ultrasuite.github.io/papers/tal_corpus_SLT2021.pdf)]