Tilawa Quran ASR assets — browser-served mirror
A re-host, not a new model. These are the unmodified v0.2.0 release assets of
yazinsai/tilawa, mirrored here for one
reason: GitHub Release downloads are served without an
Access-Control-Allow-Origin header, so a browser cannot fetch() them. Hugging
Face model repositories are CORS-enabled, which makes in-browser inference possible
without any server of our own.
Used by the in-browser build of Shazam for Quran, which identifies a recited verse fully client-side.
Contents
| File | Size | Purpose |
|---|---|---|
fastconformer_full_mixed.onnx |
88.3 MB | CTC acoustic encoder — audio → per-frame log-probabilities |
quran_ctc_tokens.json |
12.2 MB | CTC token sequences for all 6,236 verses, for re-ranking |
quran.json |
3.2 MB | Verse records (surah, ayah, Uthmani text) |
vocab.json |
21 KB | CTC token id → string |
export_metadata.json |
1 KB | Export parameters, including blank_id |
Why this architecture
The model is a CTC encoder with no autoregressive decoder. That matters for Quranic audio specifically: Whisper-family models are prone to repetition loops on repetitive, melodic, elongated speech — exactly the acoustic profile of recitation — because each token is predicted from the last. A CTC model emits per-frame probabilities independently and structurally cannot loop.
Verse identification is not plain transcription. Tilawa decodes the audio, then re-ranks all 6,236 verses by CTC log-probability, so the search is constrained to text that actually exists in the Quran rather than matching a free-form transcript.
Attribution
- SDK and export: yazinsai/tilawa (MIT), assets from release
v0.2.0 - Underlying acoustic model:
nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0(CC-BY-4.0)
Licensed CC-BY-4.0, inherited from the NVIDIA base model. No weights were retrained, quantized, or altered — this repository only changes where the bytes are served from.