🇳🇬 Stew Yanga (v0.1)
Nigeria's unified multilingual speech intelligence model.
Igbo · Yoruba · Hausa · Nigerian Pidgin · Naija English
One speech brain, five Nigerian voices. Stew Yanga (v0.1) is a unified speech intelligence system engineered for the languages of Nigeria — a market of 220+ million people that generic global models consistently under-serve. It hears Nigerian accents and languages the way Nigerians actually speak.
🧠 Capabilities
- Speech-to-Text (Ears) — real-time transcription tuned on native Nigerian speech.
- Text-to-Speech (Voice) — natural Nigerian-accented voice generation layer.
- Audio Understanding — language identification and routing across all five target languages.
- Voice Interaction — full speak-in, speak-out conversational pipeline.
- Edge & Cloud ready — lightweight checkpoints for mobile and server deployment.
🎯 Why Stew Audio
| Need | Generic global ASR | Stew Yanga (v0.1) |
|---|---|---|
| Igbo, Yoruba, Hausa support | Partial, often skipped | Core design goal |
| Nigerian accent handling | Frequent errors | Tuned on native speech |
| Pidgin awareness | Not supported | Routed voice pipeline |
| Offline / edge deployment | Heavy models | Lightweight adapters |
| Nigeria-first roadmap | No roadmap | Built for Naija, by Naija |
🚀 Quickstart
from stew_audio import StewAudio # unified interface (see files below)
model = StewAudio.load("Erog0291/stew-audio-0.1")
print(model.transcribe("my_audio.wav", language="igbo"))
Component endpoints (per-language ASR checkpoints) are published under this
namespace as stew-audio-0.1-asr-ig, stew-audio-0.1-asr-yo, stew-audio-0.1-asr-ha.
Router config: router.json.
📊 Benchmarks — Word Error Rate (%), lower is better
Evaluated on held-out native-speaker test sets (Nigerian languages, FLEURS-NG).
| Language | Generic multilingual model | Stew Yanga (v0.1) | Improvement |
|---|---|---|---|
| Igbo | 193.8 | 53.5 | 72% better |
| Yoruba | 100.4 | 61.0 | 39% better |
| Hausa | 127.2 | 38.7 | 70% better |
Measured head-to-head: same audio, same normalization, 60 native test utterances per language. A WER above 100% means the generic model produced unusable output (garbage transcripts); Stew Yanga (v0.1) cut errors dramatically across all three languages. Training: LoRA (r=16), 400 native samples/language, 240 steps, under 12 minutes on a single GPU.
Benchmark methodology: identical audio, identical text normalization, same
test speakers. Full evaluation file: eval_benchmarks.json.
🧪 Fluency evaluation
Stew Yanga (v0.1) is scored on three axes:
- Speech fluency — WER across 180 native test utterances (3 languages).
- Text fluency — transcription punctuation and diacritic fidelity.
- Understanding — language identification accuracy and code-switch handling.
⚠️ Limitations
- Version 0.1 is optimized for Igbo, Yoruba, Hausa; Pidgin routes through the English voice path.
- Heavy code-switching mid-sentence can reduce accuracy.
- Noisy far-field audio benefits from a close mic.
📄 License & Attribution
Apache-2.0 — free for commercial and research use. Built by Rejoice Gideon Emmanuel (Erogian). If Stew Audio powers your product, star the repo 🙌.