Murya — Multi-Speaker Hausa Neural Speech Synthesis
Murya is an 8-speaker Hausa text-to-speech system built on a Piper/VITS architecture and exported to ONNX for CPU-only, offline-capable inference. It runs client-side in the browser via WASM, with no GPU requirement at any stage of deployment.
The checkpoint has eight speakers (four male, four female). The public demo names two of them: Malama Asabe (female) and Malam Garba (male).
Why this model exists
At the time the WAXAL corpus was released, its authors noted that Hausa had no text-to-speech model. Murya was built to close that gap with a system that is deployable in the conditions Hausa is actually spoken in — low bandwidth, intermittent connectivity, and commodity hardware.
What distinguishes it
- Multi-speaker. Eight speakers (four male, four female) from the WAXAL Hausa TTS subset. Most publicly available Hausa TTS models are single-speaker.
- Offline-capable. ONNX/WASM client-side inference with IndexedDB caching. No Python runtime, no server round-trip, no GPU.
- Low latency. Sub-110ms time-to-first-audio-chunk on client-side WASM, with optional server fallback.
- 22,050 Hz output. The rate stored in
model.onnx.json(audio.sample_rate).
Specifications
| Architecture | Multi-speaker VITS (Piper), ONNX export |
| Speakers | 8 (4 male, 4 female). Demo names two: Malama Asabe, Malam Garba |
| Speaker map | F2=0, M3=1, M2=2, F4=3, M4=4, F1=5, M1=6, F3=7 (model.onnx.json) |
| Sampling rate | 22,050 Hz |
| Runtime | WASM ONNX client-side; FastAPI rising-edge stream server fallback |
| Fine-tuning corpus | WAXAL Hausa TTS subset (google/WaxalNLP) |
| Tonal handling | Litvinova RTL syllable-weight segmentation |
| Dialect calibration | Standard Kano Hausa (Eastern) |
Dialect scope
Acoustic tuning targets Standard Kano Hausa. Sokoto and Gobirawa acoustic variation is not currently modeled and is scheduled for a future release. Users deploying into western dialect areas should evaluate before production use.
Evaluation status
A blind, randomized Mean Opinion Score study (ITU-T P.800-style, 1–5 naturalness scale, with a real-human-speech anchor) is currently collecting ratings. No score is published until each voice clears a floor of 20 ratings. Results are not final and no MOS figure is claimed here.
Usage
Live demo and interactive evaluation: https://app.murya.ng
License
These weights are released under CC-BY-NC-SA 4.0 — free for research, academic, and other non-commercial use, with attribution, under share-alike terms.
The Murya application code is MIT and lives at https://github.com/adab-tech/hausa-ai. That license does not cover these weights.
Commercial use is not covered by this license. A separate commercial license is available, including hosted API access, self-hosted deployment, and per-language fine-tuning. Contact: hi@murya.ng
Attribution
Fine-tuned on the WAXAL corpus (CC-BY-4.0):
Diack, A., Nelson, P., Agbesi, K., et al. WAXAL: A Large-Scale Multilingual African Language Speech Corpus. arXiv:2602.02734, 2026. https://huggingface.co/datasets/google/WaxalNLP
The WAXAL corpus is used as fine-tuning data. The surrounding architecture, inference pipeline, tonal handling, and deployment stack are original work and are not derivative of the corpus.
Citation
@misc{abubakar2026murya,
title = {Murya: Multi-Speaker Offline-Capable Hausa Neural Speech Synthesis},
author = {Abubakar, Adamu Danjuma},
year = {2026},
url = {https://huggingface.co/adab-tech/murya-piper-hausa-tts}
}
© 2026 Adamu Danjuma Abubakar · https://adamu.tech