Introduction

Talqin streaming model v1.0 (Warsh) just landed on Hugging Face: talqin/talqin-stream-warsh (created 2026-11-05 ~06:20 UTC). It is a tiny causal Zipformer CTC trained from scratch to hear Warsh ‘an Nafi’ Qur’an recitation as it happens, then emit phoneme-like symbols the Talqin app maps onto ayah positions — no server required for the follow-along path.

This is not another generic language Whisper port. It is a purpose-built on-device streaming ASR for one riwayah, with a concrete 25 MB int8 artifact you can load in sherpa-onnx today.

What Shipped

  • Vendor / product: Talqin streaming model v1.0 (Warsh) — Talqin / TalqinCore
  • Modality: open-weight automatic speech recognition (streaming Qur’an / Warsh)
  • Open vs closed: weights Apache-2.0; ungated model.int8.onnx (26,484,369 bytes) + tokens.txt + SHA-256 manifest.json
  • Runnable today: sherpa-onnx OnlineRecognizer.from_zipformer2_ctc on Android, iPhone, desktop, and WebAssembly
  • Concrete fact: 22.3M parameters, 1.8% sound error on 7 held-out reciters (8.45 h); simulated home-room mic 9.1%; app phrase placement 98.3% with this model

It writes 60 sound symbols every 40 ms (~0.4 s lag), not plain text. The Talqin app places those sounds in the mushaf. Slice length is 0.32 s; claimed phone speed is 2–15× real time.

Run It

# after downloading model.int8.onnx + tokens.txt from the HF repo
pip install sherpa-onnx

import sherpa_onnx

rec = sherpa_onnx.OnlineRecognizer.from_zipformer2_ctc(
    tokens="tokens.txt",
    model="model.int8.onnx",
    num_threads=1,
)
# feed 16 kHz mono PCM in streaming chunks; read online results

Verify file hashes against manifest.json before you ship a mobile build — the app does the same check on download.

Notes for makers

  • Scope is narrow on purpose: Warsh recitation follow-along, not general Arabic ASR and not a tajweed “mistake judge” (Talqin keeps that on a larger server model)
  • Training recipe on the card: 1,511 hours across 59 reciters with noise, phone-line, and speed augmentation
  • Soft-holds unchanged (Ghana 401, Moore gated, finanfa Fon TTS NC+WIP, Eddycrack >48h, Magpie/Persian NC or gated, Underdog Woof HF 401) — Talqin cleared as a small, runnable specialty ASR with hard numbers
  • Prefer generative music/TTS when both qualify; OmniSinger is the gen pair this cycle. Ship Talqin because the artifact is primary, dated inside the window, and not a recycled low-resource Whisper mirror
  • No Gen CTA: specialty ASR does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

For makers building offline Qur’an follow-along or Warsh-aware mobile ASR, Talqin’s 25 MB int8 Zipformer is the zero-day card — pull the ONNX, hash-check the manifest, and wire it into sherpa-onnx today.

—Anabel ♡