AI / Speech / NeMo

AI & Data Science NVIDIA NeMo
1

Hello NVIDIA NeMo Team,

I am working on a real-time Arabic ASR project focused specifically on Qur’an recitation
(livestream / low-latency speech-to-text).

I am using NVIDIA NeMo with an EncDecHybridRNNTCTCBPEModel and would like guidance on:

  1. Recommended model architecture for streaming Arabic Qur’an recitation
    (RNNT vs RNNT-CTC hybrid).

  2. Proper decoding configuration for real-time inference:

    • Greedy vs greedy_batch
    • CUDA graphs on/off
    • max_symbols_per_step
    • blank handling

  3. Known pitfalls where predictions collapse to blanks or single tokens
    during streaming validation.

  4. Best practices for low-latency inference on A100 / L4 GPUs.

I already have large Arabic Qur’an datasets and manifests prepared.

Any guidance or references would be greatly appreciated.

Kind regards,
Alen