Ashuza11/afrivoices-asr-hack

Multilingual ASR for Swahili, Kikuyu, Luo, Somali, Maasai, Kalenjin — AfriVoices East Africa Hackathon # AfriVoices East Africa ASR A single unified Automatic Speech Recognition (ASR) model for six East African languages, built for the AfriVoices East Africa ASR Hackathon. ## Languages | Language | ISO Code | Dialects | |---|---|---| | Swahili | swa | Nairobi, Kisii, Wajir, Mombasa, Nakuru, Dar-es-Salaam | | Kikuyu | kik | Gĩ-Kabete, Ki-Mathira, Ki-Muranga, Ki-Ndia | | Luo (Dholuo) | luo | Nyandwat, Milambo | | Somali | som | Maxatire, Mogadishu | | Kalenjin | kln | Nandi, Kipsigis | | Maasai | mas | Kimasaai, Kisamburu | ## Approach Fine-tuning openai/whisper-small (244M parameters) on the combined multilingual dataset with: - Temperature-based language sampling (α=0.7) to balance low-resource languages - Curriculum learning: scripted speech first, then unscripted - SpecAugment + speed perturbation for data augmentation - fp16 mixed precision + gradient checkpointing on Kaggle GPU Target: CPU-only inference, ≤8 GB RAM, real-time factor ≤2x on Raspberry Pi 4. ## Hardware constraints (competition rules) - Parameters: < 1 billion - Inference: CPU-only, ≤ 8 GB RAM - Latency: RTF ≤ 2x on Raspberry Pi 4 - License: Apache-2.0 ## Repository structure ``` eval/ compute_wer.py WER/CER evaluation harness notebooks/ kaggle_01_baseline_submission.py Zero-shot baseline on Kaggle kaggle_02_data_and_finetune.py Full multilingual fine-tuning pipeline lesson2_spectrogram_wer.py Log-mel spectrogram + WER demo lesson3_ctc_vs_seq2seq.py CTC vs seq2seq latency comparison src/ train_whisper.py Whisper fine-tuning script (local smoke-test + Kaggle) ``` ## Data sources - `MCAA1-MSU/anv_data_ke` — Kikuyu, Kalenjin, Luo, Maasai, Somali (Maxatire) - `DigitalUmuganda/Afrivoice_Swahili` — Swahili - `DigitalUmuganda/Afrivoice` — Somali (Mogadishu) All datasets are licensed under CC BY 4.0. ## License Apache-2.0