Research
Notes on how the models are built and measured, written by the people who built them.
The data, in numbers
Read how it was built
- Recitation in the corpus
- 59,000+ h
- Ayat phonemized
- 6,236
- Rulings reviewed by experts
- 208
- QuranTTS, public
- 301 h
- Benchmark clips, held out
- 600
- Sep 2026 Proving the potential of ternary quantization in ASR An internal experiment: Zipformer Quran 3.1 at −1, 0 and +1 reaches 3.34% phoneme error, but it is overfit to the canonical text. Not for feedback yet; ready to build a canonicalizer on. ASR
- Sep 2026 The data behind the models: a 59,000-hour corpus, and what the models took from it Quran Lab holds more than 59,000 hours of recitation. No model trains on all of it; this is how the selections for Zipformer Quran 3 and 3.1 were made. Data
- Sep 2026 Zipformer Quran 3.1: it now writes what was said 4.13% phoneme error on all 600 benchmark clips, madd held as it was recited, and the Neural Engine. ASR
- Sep 2026 SAWT v4: 48 kHz restoration of archive recitations Bringing archive recitations back to studio quality with a 372M flow-matching generator that renders texture but never decides what was said. Restoration
- Sep 2026 Zipformer Quran 3: closing the emphatic-letter gap How ص ض ط ظ went from three times the error of plain letters to parity on every held-out set. ASR