# Amortized Neural Networks for Low-Latency Speech Recognition

arXiv: https://arxiv.org/abs/2108.01553 · HTML: https://mlsanity.com/paper/2108.01553/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 3 Aug 2021.

## Built on by (most cited first, 9 of 9)

- [Recent Advances in End-to-End Automatic Speech Recognition](https://mlsanity.com/paper/2111.01690/) (2021)
- [MeMemo: On-device Retrieval Augmentation for Private and Personalized Text Generation](https://mlsanity.com/paper/2407.01972/) (2024)
- [Personal VAD 2.0: Optimizing Personal Voice Activity Detection for On-Device Speech Recognition](https://mlsanity.com/paper/2204.03793/) (2022)
- [I3D: Transformer architectures with input-dependent dynamic depth for speech recognition](https://mlsanity.com/paper/2303.07624/) (2023)
- [Compute Cost Amortized Transformer for Streaming ASR](https://mlsanity.com/paper/2207.02393/) (2022)
- [Dual-Attention Neural Transducers for Efficient Wake Word Spotting in Speech Recognition](https://mlsanity.com/paper/2304.01905/) (2023)
- [Key Frame Mechanism For Efficient Conformer Based End-to-end Speech Recognition](https://mlsanity.com/paper/2310.14954/) (2023)
- [Improved Contextual Recognition In Automatic Speech Recognition Systems By Semantic Lattice Rescoring](https://mlsanity.com/paper/2310.09680/) (2023)
- [Accelerator-Aware Training for Transducer-Based Speech Recognition](https://mlsanity.com/paper/2305.07778/) (2023)

## Abstract (arXiv)

We introduce Amortized Neural Networks (AmNets), a compute cost- and latency-aware network architecture particularly well-suited for sequence modeling tasks. We apply AmNets to the Recurrent Neural Network Transducer (RNN-T) to reduce compute cost and latency for an automatic speech recognition (ASR) task. The AmNets RNN-T architecture enables the network to dynamically switch between encoder branches on a frame-by-frame basis. Branches are constructed with variable levels of compute cost and model capacity. Here, we achieve variable compute for two well-known candidate techniques: one using sparse pruning and the other using matrix factorization. Frame-by-frame switching is determined by an arbitrator network that requires negligible compute overhead. We present results using both architectures on LibriSpeech data and show that our proposed architecture can reduce inference cost by up to 45% and latency to nearly real-time without incurring a loss in accuracy.