saadlohani/Awaaz

Data infrastructure for speech AI in 11 low-resource languages of Pakistan: source and gate audio, annotate with native speakers, export training datasets, run ASR and TTS tuning. # Awaaz **Voice models for low resource languages.** A working pipeline for eleven low-resource Perso-Arabic languages of Pakistan and the surrounding region: find the audio, annotate it with native speakers, tune a model, hear the difference. ## Why Over 60% of the world speaks a regional language, and today's voice models don't understand them well. Pakistan is the fifth most populous country on earth and 42% of its adults cannot read, so a voice interface is often the only realistic way to reach a digital service at all. Large labs treat this as a marketing line item, not a product priority. So this project builds the whole stack itself: source audio, label it, train on it, and put a usable model in front of anyone who speaks one of these eleven languages: **Balochi · Balti · Burushaski · Hindko · Kashmiri · Khowar · Punjabi · Saraiki · Shina · Sindhi · Wakhi** ## Three apps, one pipeline | | | | ------------------- | ------------------------------------------------------------------ | | **Source Studio** | Discover audio, gate it on quality, cut it into utterances | | **Labeling Studio** | Native speakers correct machine drafts into approved transcripts | | **Frontier Tuning** | Pick a language, add audio, train a model. No ML background needed | All three read and write the same manifest (`data/manifest.db`), so a clip found in Source Studio flows straight into the labeling queue and then into a training run without anyone hand-carrying data between tools. ## Finding the audio Source Studio searches for real channels and accounts for a language, measures how much speech each one actually has (hours, video count, mean length), and queues them for a human decision: approve, reject, or preview and listen first. The right rail tracks the pipeline live: how many hours came in raw, how many survived quality gating, how many are labeled and ready to tra …