quran-whisper-base-ar-quran-executorch

Quran Whisper Base Arabic ExecuTorch for React Native

This repository contains an ExecuTorch .pte export of tarteel-ai/whisper-base-ar-quran for on-device mobile inference with react-native-executorch.

The model is architecturally compatible with openai/whisper-base and is fine-tuned for Quran recitation. The exported artifact is intended as a drop-in replacement for the Software Mansion whisper-base artifact used by react-native-executorch SpeechToText.

Files

• xnnpack/quran_whisper_base_xnnpack.pte: RN-compatible ExecuTorch model.
• tokenizer.json: Whisper tokenizer compatible with the React Native runner timestamp contract.
• export-report.json: Reproducibility report, dependency versions, validation results, hashes, and export details.
• model_processor_tokenizer.json: Tokenizer serialized directly from the Tarteel model processor for auditability.
• export_quran_whisper_executorch.py: Reproducible exporter script.

React Native Configuration

Keep EXPO_PUBLIC_HIFZ_LOCAL_MODEL_EXECUTORCH_NAME=whisper-base and multilingual mode enabled.

EXPO_PUBLIC_HIFZ_LOCAL_MODEL_EXECUTORCH_NAME=whisper-base
EXPO_PUBLIC_HIFZ_LOCAL_MODEL_URL=https://huggingface.co/ahmedsu868/quran-whisper-base-ar-quran-executorch/resolve/main/xnnpack/quran_whisper_base_xnnpack.pte
EXPO_PUBLIC_HIFZ_LOCAL_MODEL_BYTES=401540224
EXPO_PUBLIC_HIFZ_LOCAL_MODEL_SHA256=67e7c1ff62cf2c4b136f25f8f3284f2d621b536aa36093872efd71f5eb77f3df
EXPO_PUBLIC_HIFZ_LOCAL_TOKENIZER_URL=https://huggingface.co/ahmedsu868/quran-whisper-base-ar-quran-executorch/resolve/main/tokenizer.json
EXPO_PUBLIC_HIFZ_LOCAL_TOKENIZER_BYTES=2480466
EXPO_PUBLIC_HIFZ_LOCAL_TOKENIZER_SHA256=27fc476bfe7f17299480be2273fc0608e4d5a99aba2ab5dec5374b4482d1a566

Then load through:

await SpeechToTextModule.fromCustomModel(modelSource, tokenizerSource, true);

Artifact Hashes

EXPO_PUBLIC_HIFZ_LOCAL_MODEL_BYTES=401540224
EXPO_PUBLIC_HIFZ_LOCAL_MODEL_SHA256=67e7c1ff62cf2c4b136f25f8f3284f2d621b536aa36093872efd71f5eb77f3df
EXPO_PUBLIC_HIFZ_LOCAL_TOKENIZER_BYTES=2480466
EXPO_PUBLIC_HIFZ_LOCAL_TOKENIZER_SHA256=27fc476bfe7f17299480be2273fc0608e4d5a99aba2ab5dec5374b4482d1a566

Compatibility Contract

The export was validated against the known-working Software Mansion react-native-executorch-whisper-base reference artifact for [email protected].

Required methods:

• encode(waveform: Float[<=480000]) -> Float[1, 1500, 512]
• decode(tokens: Long[1, <=128], positions: Long[<=128], encoder: Float[1, 1500, 512]) -> Float[1, <=128, 51865]

Validation covered:

• Reference PTE contract inspection.
• Final PTE contract equality.
• Dummy execution for encode and decode.
• Dynamic RN-like shape checks:
  • encode: 16000, 464000, and 480000 samples.
  • decode: 1, 4, and 128 tokens.
• Tokenizer token checks for <|startoftranscript|>, <|endoftext|>, <|0.00|>, <|ar|>, and <|transcribe|>.

Tokenizer Note

The tokenizer serialized directly from tarteel-ai/whisper-base-ar-quran did not include the full timestamp-token serialization expected by the React Native runner. In particular, it omitted <|0.00|> from tokenizer.json, while the RN Whisper runner requests this token explicitly.

For the final tokenizer.json, this repo uses the Software Mansion/openai-compatible Whisper tokenizer so that timestamp IDs and the 51,865-token decoder vocabulary contract match the PTE. The original model processor tokenizer is included as model_processor_tokenizer.json for reproducibility.

Export

The export used an RN-compatible fallback path rather than accepting the current optimum-cli output. Current Optimum exports Whisper as encoder and text_decoder, while the React Native runner expects exactly encode and decode with raw-waveform preprocessing inside encode.

See export-report.json and export_quran_whisper_executorch.py for full details.

License

The base model is published as Apache-2.0 by Tarteel AI. See the upstream model page for model details.