Stargate-Hlevetela-0.77B

An Northern Languages-focused speech recognition model for research and academic use.

Overview

Stargate Africa's Hlevetela-0.77B is the Tsonga (Xitsonga) and Venda (Tshivenda) variant of the entry-level model in Stargate Africa's family of speech recognition models. Based on OpenAI's Whisper Small, it is designed to advance research into automatic speech recognition (ASR) for Tsonga (Xitsonga) and Venda (Tshivenda), with a particular focus on South African accents, speech patterns and language use.

The Stargate Africa speech recognition family aims to expand access to speech technology for South Africa's languages and communities. This initial release establishes a foundation for evaluating Venda and Tsonga speech recognition and developing more inclusive speech technologies tailored to the country's linguistic diversity.

Intended use: Research, academic study, experimentation, and evaluation.

Important: This model is intended for research and academic purposes. Commercial use requires separate authorization under the applicable model license. Please contact contact@stargateafrica.org for more capable models.

Language profile

• Language: Tsonga | Venda
• ISO 639-1 code: ts ve
• Primary region: South Africa
• Language family: Bantu
• Writing system: Latin script

Speaker coverage

Tsonga (Xitsonga) and Venda (Tshivenda) are one of South Africa's smaller official languages. Approximately 3 million and 1.6 million people speak it as their first language, making it the country's eight and tenth home language by number of speakers.

These figures describe approximate first-language speaker numbers, rather than everyone who can speak Tsonga or Venda as an additional language.

Model details

Evaluation and benchmarks

The model should be evaluated against appropriate Venda and Tsonga speech datasets to establish its recognition accuracy and limitations.

Benchmark results

Control model: ggml-whisper-medium.bin Test model: ggml-Hlevetela-0.77B.bin

The benchmark dataset is comprised of a 50/50 split between DSFSI and NCHLT test splits.

Inference example

For help with PyWhisperCPP please visit https://github.com/absadiki/pywhispercpp

from pathlib import Path

from huggingface_hub import hf_hub_download
from pywhispercpp.model import Model


def main() -> None:
    # Download the model from Hugging Face
    model_path = hf_hub_download(
        repo_id="StargateAfrica/Hlevetela-0.77B",
        filename="ggml-Hlevetela-0.77B.bin",
    )

    # Validate the input audio
    audio_path = Path("ve.wav")
    if not audio_path.is_file():
        raise FileNotFoundError(
            f"Audio file not found: {audio_path.resolve()}"
        )

    # Load the Whisper-compatible model
    model = Model(
        model_path,
        language="ve",
        n_threads=8,
        no_timestamps=True,
        single_segment=True,
    )

    # Transcribe the audio
    segments = model.transcribe(
        str(audio_path),
        language="ve",
        n_threads=8,
    )

    # Print the combined transcription
    transcription = " ".join(
        segment.text.strip()
        for segment in segments
        if getattr(segment, "text", "").strip()
    ).strip()

    print(transcription)


if __name__ == "__main__":
    main()

Intended use and limitations

Hlevetela-0.77B is intended to support research into Tsonga or Venda ASR and the development of speech technologies for South African users.

Potential research applications include:

• Tsonga and Venda speech recognition and transcription.
• Comparative evaluation of ASR models.
• Research into regional accents and speech variation.
• Development of language technologies for underrepresented communities.
• Academic research into multilingual and locally adapted speech systems.

Recognition performance may vary with accent, dialect, speaking rate, background noise, recording quality, code-switching and domain-specific vocabulary. Users should independently evaluate the model for their intended application.

This release should not be assumed to provide production-grade accuracy or to perform equally well across all Tsonga or Venda-speaking communities.

License and usage

This model is released under apache-2.0 (apache-2.0).

The license governs permitted use, redistribution and other applicable conditions. Not intented for commercial use.

Please consult the license terms before using, adapting or redistributing the model.

Acknowledgements and references

• Statistics South Africa — Mbalo Brief, November 2023
• Xitsonga
• Tshivenda
• OpenAI Whisper