Ahmad ZushanAI · Systems · Platform

Home/2026 builds/№ 39

39 · Voice & language

Urdu TTS dataset — 28 hours of aligned speech

Sentences from the Common Voice collection, Tatoeba and Urdu Wikipedia, cleaned and bucketed by length; a resumable synthesiser running 36 workers; post-processing to trimmed, loudness-normalised 22.05, 24 and 48 kHz sets with LJSpeech and multi-speaker metadata.

PeriodOct 2026StatusDatasetAreaVoice & language

Highlights

  1. 28.3 hours with zero dropped clips for about $23 of API time, measured at roughly 800 clips and 57k billable characters per hour of audio.
  2. A 9.85 GB hand-off kit with a verify script and a roadmap for the training machine: checks, a single-speaker Piper voice, a four-speaker model, an optional F5 run.
  3. A 16 kHz subset exported in the Bol Trainer format as a minority supplement for speech recognition.