DEVBLOCK-TECHNOLOGIES-LIMITED/eti-yoruba-asr
Domain:
natural language processingRecord type:
model
Creator:
DEV
Host:
Ẹtí — open-source Yoruba ASR (Whisper-small LoRA fine-tune) by DevBlock Technology Limited
# Ẹtí — Open Yoruba ASR
**Ẹtí** (Yoruba: *ẹtí*, "ear") is an open-source
Whisper-small model fine-tuned
for **Yoruba automatic speech recognition**, published by
**DevBlock Technology Limited** (devblocktechnologies.com).
Yoruba (~50M speakers) has limited commercial ASR coverage. Ẹtí is an openly
licensed baseline — including a CPU-friendly CTranslate2 build for edge and
low-cost serving.
---
## Model summary
| | |
|---|---|
| **Base** | `openai/whisper-small` (MIT) |
| **Adaptation** | LoRA on `q_proj`/`v_proj` (r=16, α=32), merged |
| **Language** | Yoruba (`yo`) |
| **Formats** | Transformers + CTranslate2 / faster-whisper |
| **License** | MIT |
**Model (Hugging Face):** `devblockHQ/eti-yoruba-asr`
## Evaluation (WER, held-out, lower = better)
| Set | n | Raw | No-diacritics |
|---|---|---|---|
| Read speech (base-corpus test) | 40 | 0.557 | 0.473 |
| Conversational (`thisniyi/yoruba-speech-project-v2`) | 50 | 1.087 | 0.822 |
Conversational Yoruba is far harder than read/news speech — the gap is the
main signal for where the model needs more conversational training data.
Production target (not yet met): WER ≤ 0.20 API / ≤ 0.28 telephony.
## Quickstart
Transformers:
```python
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa
repo = "devblockHQ/eti-yoruba-asr"
model = WhisperForConditionalGeneration.from_pretrained(repo)
proc = WhisperProcessor.from_pretrained(repo, language="yoruba", task="transcribe")
audio, sr = librosa.load("clip.wav", sr=16000, mono=True)
feats = proc(audio=audio, sampling_rate=16000, return_tensors="pt").input_features
print(proc.batch_decode(model.generate(feats, language="yoruba", task="transcribe"),
skip_special_tokens=True)[0])
```
faster-whisper (CPU-friendly CT2):
```python
from faster_whisper import WhisperModel
m = WhisperModel("devblockHQ/eti-yoruba-asr/ct2", device="cpu", compute_type="int8")
print(" ".join(s.text for s in m.transcribe("clip.wav", language="yo", beam_size=5 …