README.md

---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: file_name
    dtype:
      audio:
        sampling_rate: 24000
  - name: speaker_id
    dtype: int64
  - name: ref_trans_clean
    dtype: string
  - name: asr_trans
    dtype: string
  - name: cer
    dtype: float64
  - name: age
    dtype: string
  - name: gender
    dtype: string
  - name: raw_stm_content
    dtype: string
  splits:
  - name: train
    num_bytes: 54568714206
    num_examples: 51295
  download_size: 54968584230
  dataset_size: 54568714206
license: cc-by-4.0
task_categories:
- text-to-speech
language:
- pt
---


# FalAR-TTS

[![Hugging Face Dataset](https://img.shields.io/badge/huggingface-dataset-blue)](https://huggingface.co/datasets/inesc-id/FalAR-TTS)

FalAR-TTS is a subset of the FalAR dataset (https://huggingface.co/datasets/inesc-id/FalAR) tailored for speech synthesis in **European Portuguese**. To minimize longitudinal variations in the acoustic characteristics of the speech signals, the subset is restricted to utterances recorded between 2016 and 2017. Furthermore, to ensure highly reliable timestamped speech alignments, an additional filtering criterion was applied: recordings were only selected if they exceeded 5 seconds in duration and exhibited a reference-to-automatic alignment character error rate (CER) below 15%.

# Content Description
The resulting FalAR-TTS dataset comprises 280 speakers, consisting of 180 male speakers (accounting for 221 hours of audio) and 100 female speakers (97
hours). All audio files are delivered at a sampling rate of 24 kHz.

## Dataset field description
### file_name
The actual audio data for the segment.
### speaker_id
Integer ID assigned to the speaker recorded in this file during the annotation process.
### gender
Gender of the speaker recorded in this file: Male (M) vs Female (F)
### age
The age of the speaker (in years) at the time of the recording.
### ref_trans_clean
A clean orthographic transcript sourced from the Portuguese Parliament archive, which is natively devoid of
repetitions, hesitations, and speech disfluencies.
### asr_trans
Automatic content transcription produced by CAMOES ASR models (https://huggingface.co/datasets/inesc-id/camoes_asr).
### cer
Character Error Rate computed from the alignment between the reference (ref_trans_clean) and the automatic transcripts (asr_trans). Values between 0 and 1.
### raw_stm_content
A fine-grained, timestamped automatic transcript of the spoken content.

## Segment Time Mark Annotations
To expand the characterization of the dataset, each audio file — nominally containing speech from a single individual — was processed using the Pyannote speaker diarization framework.
The resulting outputs were utilized to isolate overlapping speech intervals and to extract speaker embeddings via SpeechBrain’s pre-trained ECAPA-TDNN model. These embeddings allowed for the detection of file-wise acoustic outliers
based on average pairwise cosine similarity calculations.
In addition to the raw audio, the dataset includes a "raw_stm_content" field, which enables the reconstruction of a standard Segment Time Mark (STM)
file. This file contains timestamped speech segments and their corresponding automatic transcriptions, augmented with two distinct metadata annotations
appended in an [A|B] format:
– A: The average cosine similarity between the current speaker embedding and
those computed for the remaining speaker turns within the file (evaluates to
"n/a" if the speaker segmentation yields only a single speaker turn);
– B: The ratio of the overlapping speech span to the total segment duration

Interpretation of A values: each audio file is assumed to contain speech from a single speaker. However, if the diarization tool (e.g., Pyannote) segments the audio into multiple speaker turns, an ECAPA-TDNN embedding is extracted for each segment. If a given pseudo-speaker is acoustically consistent with the remaining pseudo-speakers identified within that file, the average cosine similarity between its embedding and the others will be high; a low similarity, conversely, indicates the likely presence of a different speaker.

## License

This dataset is listed on Hugging Face under *CC-BY-4.0*.