dev-ahmedhany/stt_ar_fastconformer_hybrid_large_streaming_pcd_v1.0-sherpa

stt_ar_fastconformer_hybrid_large_streaming_pcd_v1.0 — sherpa-onnx 3-file split

Streaming-trained Arabic FastConformer-Hybrid-Large (RNNT + CTC heads), exported for sherpa-onnx's OnlineRecognizer with cache-aware encoder state.

Geometry

• att_context_size: [70, 13]
• Left context: 5600 ms of past audio (memory window, no extra latency)
• Right context (lookahead): 1040 ms — constant emission delay
• Encoder subsampling: 8× (10 ms frames → 80 ms output steps)

Files

file	purpose
encoder.onnx	streaming encoder with cache I/O slots
decoder.onnx	RNNT prediction network
joiner.onnx	joint network
tokens.txt	SentencePiece vocab (one token per line, token\tid)

Usage (Dart / sherpa-onnx)

final transducer = sherpa.OnlineTransducerModelConfig(
  encoder: 'encoder.onnx',
  decoder: 'decoder.onnx',
  joiner:  'joiner.onnx',
);
final model = sherpa.OnlineModelConfig(
  transducer: transducer,
  tokens: 'tokens.txt',
  modelType: 'transducer',
  provider: 'cpu',
);

Provenance

• Base: nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0
• Fine-tuned with full FT (no LoRA — masks rejected adapter-only training)
• Data: tarteel-ai/everyayah, tartil reciters only, 100h cap
• License: CC-BY-4.0 (inherits NVIDIA pcd license)