# AI / Speech / NeMo

**URL:** <https://forums.developer.nvidia.com/t/ai-speech-nemo/355034>\
**Category:** NVIDIA NeMo\
**Tags:** nemo\
**Created:** [December 18, 2025, 12:45am UTC](https://forums.developer.nvidia.com/t/ai-speech-nemo/355034 "2025-12-18T00:45:49Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![kontakt109](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@kontakt109](https://forums.developer.nvidia.com/u/kontakt109)\
**Post date:** [December 18, 2025, 12:45am UTC](https://forums.developer.nvidia.com/t/ai-speech-nemo/355034/1 "2025-12-18T00:45:49Z")

</div>

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:

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