Speech Resources in the Tamasheq Language

Marcely Zanon Boito, Fethi Bougares, Florentin Barbier, Souhir Gahbiche, Loïc Barrault, Mickael Rouvier, Yannick EstèvearXiv 2022 · arXiv

TL;DR

Researchers have released two significant new datasets for Tamasheq, a low-resource Berber language spoken in the Sahel. The release includes a 17-hour parallel corpus of Tamasheq audio translated into French and a massive 671-hour unlabeled multilingual audio collection, providing a vital foundation for developing speech translation and self-supervised models in under-represented languages.

What Problem Does This Paper Address?

The "digital divide" in AI is stark: while high-resource languages like English or Mandarin have massive datasets, thousands of other languages lack the basic building blocks for modern speech technology. Tamasheq, spoken by approximately 500,000 people across Mali and Niger, is one such "developing language."

Current speech processing pipelines are often brittle when applied to diverse, smaller datasets. The authors address this by providing the community with high-quality, real-world data sourced from radio broadcasts, specifically designed for the IWSLT 2022 low-resource speech translation challenge.

Methods

The researchers partnered with the Fondation Hirondelle, a Swiss non-profit that supports independent media in conflict zones. They web-crawled daily news broadcasts from Studio Kalangou (Niger) and Studio Tamani (Mali).

Key technical steps included:

  • Translation: They commissioned native Tamasheq speakers to translate audio directly into French text (skipping the transcription phase to save resources).
  • Segmentation: They used the LIUM_SpkDiarization tool to automatically slice long radio episodes into "breath turns"—shorter, manageable audio chunks suitable for machine learning models.
  • Post-processing: They applied standard normalization and tokenization for the French text and filtered out non-speech segments like music and jingles.
  • Baseline Modeling: To set a benchmark, they trained an end-to-end speech-to-text Transformer model using the fairseq S2T toolkit.

Datasets

The paper introduces two primary resources:

  1. Tamasheq-French Parallel Corpus: A targeted dataset for speech translation. It comes in two flavors: a "clean" version (17 hours) and a "full" version (19 hours) that includes noisier segments.
  2. Niger-Mali Audio Collection: A massive 671-hour unlabeled corpus (641 hours after segmentation) featuring five languages spoken in the same geographic region: French, Fulfulde, Hausa, Tamasheq, and Zarma.
Table 1: Statistics for the clean (left) and full (right) Tamasheq-French parallel corpus.

Key Insights and Ideas

  • Radio as a Resource: Broadcast news is a goldmine for low-resource languages. It provides high-quality audio, professional speakers, and diverse topics (interviews, news, ads).
  • The "More is Better" Rule: Interestingly, the baseline model performed slightly better on the "full" 19-hour dataset than the "clean" 17-hour one. This suggests that in extreme data-scarcity scenarios, the benefit of having more data outweighs the drawback of that data being potentially noisy.
  • Geographic Transfer Learning: By including other languages from the same region (like Hausa and Zarma), the authors enable research into whether models can learn useful phonetic or acoustic features from neighboring languages to improve performance on the target language.
  • Gender Bias Warning: The authors honestly report a significant gender imbalance—the vast majority of the broadcast data features male speakers. This is a crucial metadata point for researchers to ensure their models don't inherit this bias.

Key Results

The baseline results highlight the difficulty of the task. Using only the parallel data, the models achieved BLEU scores between 1.8 and 2.3.

While these numbers are low compared to high-resource benchmarks, they serve as a realistic starting point. The authors argue that these scores demonstrate that parallel data alone isn't enough; future success will likely require leveraging the larger unlabeled corpus through self-supervised pre-training (e.g., using models like wav2vec 2.0 or HuBERT).

Limitations and Future Work

  • Gender Imbalance: The dataset is heavily skewed toward male voices, reflecting the demographics of the radio hosts and interviewees.
  • Transcription Gap: The parallel corpus lacks Tamasheq transcriptions (it only has French translations), which prevents the training of traditional ASR (Automatic Speech Recognition) models.
  • Future Directions: The authors suggest that the next logical step is to use the 641-hour unlabeled collection for self-supervised fine-tuning. They also propose incorporating pre-trained French language models (like CamemBERT) to bolster the translation decoder.

Summary updated

Abstract

In this paper we present two datasets for Tamasheq, a developing language mainly spoken in Mali and Niger. These two datasets were made available for the IWSLT 2022 low-resource speech translation track, and they consist of collections of radio recordings from daily broadcast news in Niger (Studio Kalangou) and Mali (Studio Tamani). We share (i) a massive amount of unlabeled audio data (671 hours) in five languages: French from Niger, Fulfulde, Hausa, Tamasheq and Zarma, and (ii) a smaller 17 hours parallel corpus of audio recordings in Tamasheq, with utterance-level translations in the French language. All this data is shared under the Creative Commons BY-NC-ND 3.0 license. We hope these resources will inspire the speech community to develop and benchmark models using the Tamasheq language.
BibTeX
@misc{boito2022speech,
  author = {Marcely Zanon Boito and Fethi Bougares and Florentin Barbier and Souhir Gahbiche and Loïc Barrault and Mickael Rouvier and Yannick Estève},
  title = {{Speech Resources in the Tamasheq Language}},
  year = {2022},
  eprint = {2201.05051},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2201.05051},
}
Figure