cdli/whisper-small_finetuned_ghanian_ga_nonstandard_speech_v2.0

Whisper small for Ga, stage 2 (nonstandard speech)

This is the small Whisper model (about 242M parameters, openai/whisper-small), fine-tuned for Ga speech recognition.

This is stage 2 of a two-stage fine-tuning. It starts from the stage 1 model, ElizabethMwangi/ga-standard-stage1, and is trained further on nonstandard speech.

Training

• Starting model: ElizabethMwangi/ga-standard-stage1
• Data: cdli/ghanian_ga_nonstandard_speech_v1.0
• 1000 steps, learning rate 5e-5, batch size 8 with 4 accumulation steps (effective batch 32)
• Augmentation: added noise, volume change and speed change (no SpecAugment)

Ga is not a Whisper language, so the Yoruba language token (yo) was used during training. Set language="yo" when transcribing.