Haramiday

owe-voice-whisper-small-yoruba

README

License: apache-2.0

Results

Table with columns: Test set, Metric, This model, Base model (zero-shot)
Test set	Metric	This model	Base model (zero-shot)
Òwe-Voice proverbs, 3 unseen speakers (750 utt.)	WER	0.458 [0.443, 0.474]	0.863
	CER	0.159	0.381
	WER ignoring all diacritics	0.361	0.666
	Tone error rate (on correctly recognised vowels)	0.153	0.457
FLEURS Yorùbá test (conventional read speech)	WER	0.788	0.702

Seed variability on the proverb test set: 0.462 ± 0.003 over 3 seeds.

Scoring. References and hypotheses are normalised identically: Unicode NFC, lower-casing, removal of .,!?;:"“”‘’()[]/\, whitespace collapsed; tone marks and underdots are kept. WER/CER are corpus-level (total edits / total reference units) with 95% bootstrap confidence intervals over utterances. FLEURS = Yorùbá (yo_ng) test set, 817 utterances ≤ 30 s.

Intended use and limitations

Intended for research on Yorùbá ASR, especially proverbial / culturally grounded speech.
Domain-specialised. Fine-tuning on proverbs changes performance on conventional Yorùbá (FLEURS WER 0.702 → 0.788). For general-domain Yorùbá, evaluate before use.
Trained and tested on read studio speech; the test speakers are 3 adult female speakers. Conversational, noisy or dialectal speech is not covered.
Outputs follow standard Yorùbá orthography as used in the corpus, which is not fully consistent in word segmentation (e.g. the progressive marker ń written attached or separate).

Usage

python

import torch, librosafrom transformers import WhisperProcessor, WhisperForConditionalGenerationproc = WhisperProcessor.from_pretrained("Haramiday/owe-voice-whisper-small-yoruba", language="yoruba", task="transcribe")model = WhisperForConditionalGeneration.from_pretrained("Haramiday/owe-voice-whisper-small-yoruba").eval()audio, _ = librosa.load("proverb.wav", sr=16000)features = proc.feature_extractor(audio, sampling_rate=16000, return_tensors="pt").input_featureswith torch.no_grad():    ids = model.generate(features, language="yoruba", task="transcribe")print(proc.tokenizer.batch_decode(ids, skip_special_tokens=True)[0]) # Convenience alternative (decoding can differ slightly from the reported numbers):# from transformers import pipeline# asr = pipeline("automatic-speech-recognition", model="Haramiday/owe-voice-whisper-small-yoruba")# asr("proverb.wav", generate_kwargs={"language": "yoruba", "task": "transcribe"})

Audio must be 16 kHz mono.

Training

Training data. Òwe-Voice (Abolade, 2026): read Yorùbá proverbs recorded in a studio by 12 volunteer native speakers. The corpus is not publicly released; access is at the discretion of its creators. Fine-tuning used a speaker-disjoint split: 8 speakers / 3,996 utterances / 5.17 h for training, 1 speaker / 250 utterances for checkpoint selection (dev), and 3 unseen speakers / 750 utterances / 0.95 h for testing. Each proverb was read exactly once.

Procedure. Full fine-tuning, learning rate 1e-5, linear warm-up 200 steps, batch size 16, fp16, gradient checkpointing, up to 10 epochs; the checkpoint with the lowest dev WER is kept. Language and task fixed to Yorùbá transcription. Seed 42.

Hardware: 1 × NVIDIA H200. Software: transformers 4.57.1, torch 2.8.

Citation

bibtex

@inproceedings{owevoice2026speech,  title     = {Evaluating and Fine-Tuning Speech Technologies on Culturally Grounded Yor\`ub\'a Proverbs Using the \`Owe-Voice Proverb Speech Corpus},  author    = {Fagoroye, Ayomide Emmanuel and Odumuyiwa, Victor and Abolade, Daud and Ola-obaado, Similoluwa and Taiwo, Iyanda},  booktitle = {Africa in AI Affinity Group at Neural Information Processing Systems (NeurIPS)},  year      = {2026}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Haramiday

Model Tree

Base

steja/whisper-small-yoruba

Fine-tuned

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality