On the compression of shallow non-causal ASR models using knowledge distillation and tied-and-reduced decoder for low-latency on-device speech recognition.
Nagaraj AdigaJinhwan ParkChintigari Shiva KumarShatrughan SinghKyungmin LeeChanwoo KimDhananjaya GowdaPublished in: CoRR (2023)
Keyphrases
- data warehouse
- speech recognition
- automatic speech recognition
- low latency
- language model
- speech recognizer
- hidden markov models
- acoustic models
- speech synthesis
- pattern recognition
- noisy environments
- speech signal
- speech recognition technology
- speech retrieval
- speech recognition systems
- word error rate
- speaker independent
- speaker identification
- computer vision
- probabilistic model
- real time
- speech recognizers
- stream processing
- virtual machine
- low complexity
- high throughput