pdTetteh/fedspeech-ghana

Federated ASR baseline across Ghanaian accent splits using wav2vec2, with per-client evaluation and reproducible experiments. ## Federated Baseline (FedAvg Across Accents) After establishing a centralized baseline, we simulate a simple federated learning (FL) setup where each accent is treated as a client: - Clients: Akan, Ga, Ewe, Dagomba - Server: maintains a global wav2vec2 model - Communication rounds: 3 - Algorithm: FedAvg - Local objective: CTC loss on each client’s local data - Privacy: none yet (this is a non-private baseline) ### Federated training setup We start from the same pretrained model as in the centralized baseline: - Base model: facebook/wav2vec2-base-960h - Trainable parameters: only the CTC head (lm_head); the wav2vec2 encoder is frozen for stability on this small dataset. - Optimizer: AdamW - Local training per round/client: - local_epochs = 1 - lr = 1e-5 - batch_size = 2 - max_batches = 5 (≈10 examples/client/round) - Aggregation: FedAvg weighted by number of local examples This is intentionally conservative: each client only makes a very small gradient step per round, so the global model stays close to the strong pretrained initialization. ### Results: centralized vs. federated All numbers are WER (↓) on the held-out test split, computed with jiwer.wer using text_norm as reference. #### Overall and per-accent WER | Model / Setting | Test WER | Akan | Dagomba | Ewe | Ga | |------------------------------------|---------:|------:|--------:|------:|------:| | Centralized baseline (zero-shot) | 0.4605 | 0.5180 | 0.4248 | 0.4395 | 0.4299 | | Federated (FedAvg, 3 rounds, safe) | 0.4607 | 0.5186 | 0.4243 | 0.4395 | 0.4315 | Up to normal sampling noise, the global model after 3 rounds of FedAvg matches the centralized baseline: - Overall WER remains ≈0.46. - Accent-specific WERs differ by ≤0.01 in absolute terms. In other words, with small local updates on a tiny corpus and a frozen encoder, federated fine-tuning neither helps nor hurts performance: the global model essentially stays at the pretrained optimum. ### Interpretation …