YazoPi/bambara-bomu-tts-dataset

Bambara-Bomu Dataset — Autoregressive TTS

Description

This dataset is dedicated to text-to-speech (TTS) synthesis in Bambara (bm) and Bomu (bmq) using an autoregressive approach.

It was built by combining several audio and text sources in Bambara and Bomu, then encoded to form aligned (text, audio) pairs.

Data Sources

This dataset is the result of merging the following three datasets:

Dataset	Link
Panga-Azazia/boomu-data	HuggingFace
RobotsMali/kunkado	HuggingFace

Construction Pipeline

1. Audio Encoding — SNAC (Multi-Scale Neural Audio Codec)

Audio files from the source datasets were encoded with SNAC, a multi-scale neural audio codec. SNAC produces a hierarchical representation across 3 resolution levels (coarse → fine), yielding 7 tokens per audio frame.

These raw tokens were then flattened into a 1D sequence through hierarchical packing: each of the 7 slots per frame receives a unique offset in the LLM vocabulary, ensuring that no two tokens from different levels share the same ID.

2. Text Encoding — YazoPi/Anu

Bambara and Bomu text was tokenized using the tokenizer of the YazoPi/Anu model, producing the text_tokens sequences.

3. Autoregressive Task Construction

Text and audio tokens were concatenated to form an autoregressive text-to-speech modeling task, where the model learns to predict audio tokens conditioned on the input text.

License

Non-commercial use due to some data.