HyperConformer: Multi-head HyperMixer for Efficient Speech Recognition

HyperConformer: Multi-head HyperMixer for Efficient Speech Recognition

Paper in proceedings (conference paper)

State-of-the-art ASR systems have achieved promising results by modeling local and global interactions separately. While the former can be computed efficiently, global interactions are usu- ally modeled via attention mechanisms, which are expensive for long input sequences. Here, we address this by extending Hy- perMixer, an efficient alternative to attention exhibiting linear complexity, to the Conformer architecture for speech recogni- tion, leading to HyperConformer. In particular, multi-head Hy- perConformer achieves comparable or higher recognition per- formance while being more efficient than Conformer in terms of inference speed, memory, parameter count, and available train- ing data. HyperConformer achieves a word error rate of 2.9% on LibriSpeech test-clean with less than 8M neural parameters and a peak memory during training of 5.7GB, hence trainable with accessible hardware. Encoder speed is between 38% on mid-length speech and 56% on long speech faster than an equiv- alent Conformer.1)

State-of-the-art ASR systems have achieved promising results by modeling local and global interactions separately. While the former can be computed efficiently, global interactions are usu- ally modeled via attention mechanisms, which are expensive for long input sequences. Here, we address this by extending Hy- perMixer, an efficient alternative to attention exhibiting linear complexity, to the Conformer architecture for speech recogni- tion, leading to HyperConformer. In particular, multi-head Hy- perConformer achieves comparable or higher recognition per- formance while being more efficient than Conformer in terms of inference speed, memory, parameter count, and available train- ing data. HyperConformer achieves a word error rate of 2.9% on LibriSpeech test-clean with less than 8M neural parameters and a peak memory during training of 5.7GB, hence trainable with accessible hardware. Encoder speed is between 38% on mid-length speech and 56% on long speech faster than an equiv- alent Conformer.1)

Hypernetworks, HyperMixer, Efficient Auto- matic Speech Recognition, LibriSpeech, SpeechBrain

Hypernetworks, HyperMixer, Efficient Auto- matic Speech Recognition, LibriSpeech, SpeechBrain

MAI, F.; ZULUAGA-GOMEZ, J.; PARCOLLET, T.; MOTLÍČEK, P.

2024

20.08.2023

International Speech Communication Association

Dublin

Proceedings of the Annual Conference of International Speech Communication Association, INTERSPEECH

1990-9772

Proceedings of Interspeech

2023

08

French Republic

2213

2217

5

https://www.isca-archive.org/interspeech_2023/mai23_interspeech.pdf

mai23_interspeech