Accuracy issues with RIVA asr

AI & Data Science Deep Learning (Training & Inference) Riva
1

Hardware - AGX Orin
Hardware - CPU
Operating System：Ubuntu 20.04
Riva Version ：2.8.1

Currently, when using a Mandarin model with a microphone, the following issues may arise:

  1. Poor accuracy in single-person daily conversations.
  2. Unsatisfactory results in multi-person conversations for speech recognition.
  3. Word boosting, such as for specific terms like company names, does not improve accuracy. Is it possible that provide a C++ example?

Thank you for your help.

2

HI @u9713112

Thanks for your interest in Riva,

Kindly request to provide more information

  1. Poor accuracy in single-person daily conversations - Can you record and share the voice samples used anlong with accuracy values

  2. Unsatisfactory results in multi-person conversations for speech recognition. - Can you record and share the voice sample and provide more details on what is unsatisfactory in the results

    1. Word boosting, such as for specific terms like company names, does not improve accuracy. Is it possible that provide a C++ example?

Please find the word boosting example

riva_streaming_asr_client --audio_file=audio_file.wav --boosted_words_file=boosted_words_file.txt --boosted_words_score=100. --interim_results=false

boosted_words_file.txt should contain the list of words that need to be boosted (one word on each line)