whisper-small-yoruba-finetuned

Whisper Small Fine-Tuned — Yoruba

This model is a fine-tuned version of openai/whisper-small for Yoruba speech recognition. It was trained on Common Voice 25.0 (yo) augmented with Google Fleurs (yo_ng) for a total of 3,718 training samples.

WER (raw, diacritics preserved):

• Test set (1,071 samples) : 73.67 %
• Validation set (975 samples) : 71.50 %

The WER is computed without any text normalisation in order to preserve Yoruba diacritics (tone marks, sub-dots, etc.). Lowercasing or stripping diacritics would inflate the reported score, but the model genuinely generates them.

Model description

Whisper Small is a transformer-based encoder-decoder model for automatic speech recognition (ASR), pre-trained on 680k hours of multilingual data. It uses 80-channel log-mel spectrograms as input and outputs a sequence of text tokens via autoregressive decoding with cross-attention to the encoder.

This fine-tuned version adapts the model specifically to Yoruba, a Niger-Congo language spoken in Nigeria, Benin and Togo. The model generates Yoruba text with full diacritic marks (à, è, ẹ, í, ò, ọ, ú, ṣ, etc.).

Intended uses & limitations

Intended uses:

• Transcribing Yoruba speech from audio recordings
• ASR research for low-resource African languages
• Benchmarking diacritic-preserving ASR systems

Limitations:

• Trained on ~4.3 epochs only (500 steps). More training may improve WER.
• Training data is mostly read-speech from Common Voice; performance on conversational/spontaneous Yoruba may be lower.
• The model retains Whisper's 30-second audio window.
• Diacritic coverage may be imperfect for some Yoruba dialects.

Training and evaluation data

• Common Voice 25.0 (yo) — 1,422 train / 975 validation / 1,071 test samples
• Google Fleurs (yo_ng, raw_transcription field to preserve diacritics) — 2,339 train samples (+43 removed by the >30 s filter)

Only clips shorter than 30 seconds were kept. Combined training set: 3,718 samples.

Training procedure

Hardware

• 2 × NVIDIA Tesla T4 (15.6 GB each)
• Distributed Data Parallel (DDP) via accelerate launch
• Mixed precision: fp16 (Native AMP)
• Total training time: ~26 minutes (1,605 seconds)

Training hyperparameters

• learning_rate: 1e-05
• train_batch_size (per device): 16
• eval_batch_size (per device): 16
• gradient_accumulation_steps: 1
• total_train_batch_size (effective): 32
• total_eval_batch_size (effective): 32
• seed: 42
• optimizer: AdamW (torch fused) with betas=(0.9, 0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 50
• max_steps: 500
• gradient_checkpointing: True
• eval_strategy: steps (every 500 steps)
• save_strategy: steps (every 500 steps, max 2 checkpoints)
• generation_max_length: 225
• predict_with_generate: True
• mixed_precision_training: Native AMP (fp16)

Training results

Qualitative examples (test set, seed=42)

Framework versions

• Transformers 5.12.1
• PyTorch 2.10.0+cu128
• Datasets 5.0.0
• Tokenizers 0.22.2
• Accelerate 1.14.0