Seeking Guidance on Custom Urdu ASR Training Data and Vocabulary Expansion

[Shaukataliii]

Hello,
I am a developer working on a project involving the development of an Urdu Automatic Speech Recognition (ASR) system using the Kaldi ASR toolkit. I am encountering two specific challenges and would greatly appreciate your insights.

Challenges

1. Acquiring Transcriptions for Custom Urdu Dataset:

• Issue: Obtaining accurate transcriptions for a substantial custom Urdu language dataset, tailored for industry-specific use, has proven challenging.

• Request: Seeking guidance or suggestions on cost-effective solutions or resources that could assist in obtaining accurate transcriptions.

2. Optimizing Kaldi ASR for Recognizing Unseen Words:
   • Issue: We aim to optimize the Kaldi ASR model to efficiently recognize new words it may encounter during inference, especially industry-specific jargon.
   • Request: Looking for insights or recommendations on approaches to handle previously unseen words and enhance the model's adaptability.

Thank you for your time and consideration.

[judyfong]

For two I recommend looking at the icelandic althingi recipe a bit. https://github.com/cadia-lvl/althingi-asr We use sub word modeling and also fst and regular expressions through Thrax. We're looking to merge the recipe into kaldi-asr soon.