alimaslax/somali-duplex-plotly

# Somali Duplex Dashboard Somali Duplex is a Somali text-to-speech research and data pipeline. Its goal is to build a model that learns Somali pronunciation, phrasing, prosody, and natural speaker characteristics from authorised native Somali recordings paired with accurate Somali transcripts. This repository is the Plotly Dash interface for exploring that work. The dashboard is intended to make the corpus and training journey legible: what audio has been collected, where it is in the processing pipeline, how much verified material is available, and how model experiments perform. ## Project goal The target is not English speech delivered with a Somali accent. It is a model that can turn written Somali into clear, natural Somali speech. The project uses a pretrained multilingual speech model as a starting point, then adapts it with clean Somali speech and transcript pairs. A corpus of roughly 100 hours is a useful first target; larger, well-reviewed corpora support stronger quality and broader voice coverage. ## End-to-end pipeline ```text Authorised Somali recordings | v Source collection and rights review | v Silero VAD segmentation (20–30 second speech clips) | v DeepFilterNet 3 denoising and loudness normalisation | v Mono 24 kHz, 16-bit FLAC + metadata | v Transcription, timestamp review, and speaker/quality checks | v Optional punctuation recovery from real word pauses | v Versioned train / development / test datasets | v CosyVoice 3 fine-tuning and held-out evaluation ``` ### 1. Collection and preparation Recordings must be Somali-language material that the project is authorised to process and train on. The preferred sources are clean, intelligible speech with reliable transcripts: studio podcasts, narration, and formal broadcast speech are useful complements when their speakers and usage rights are understood. Raw recordings are segmented with Silero VAD, then processed in one local pass with DeepFilterNet 3 and loudness normalisation. Final data is …