GPU-Accelerated Viterbi Exact Lattice Decoder for Batched Online and Offline Speech Recognition

TL;DR

This paper introduces a GPU-accelerated Viterbi lattice decoder that significantly improves speed and memory efficiency for speech recognition, enabling real-time streaming and large graph processing on diverse hardware.

Contribution

The paper presents a novel GPU-based Viterbi decoder with optimized memory, I/O, and parallelism, outperforming existing decoders in speed and scalability for speech recognition tasks.

Findings

Up to 240x speedup over single-core CPU decoding

Up to 40x faster than current state-of-the-art GPU decoders

Supports larger graphs and multiple streams efficiently

Abstract

We present an optimized weighted finite-state transducer (WFST) decoder capable of online streaming and offline batch processing of audio using Graphics Processing Units (GPUs). The decoder is efficient in memory utilization, input/output (I/O) bandwidth, and uses a novel Viterbi implementation designed to maximize parallelism. The reduced memory footprint allows the decoder to process significantly larger graphs than previously possible, while optimizing I/O increases the number of simultaneous streams supported. GPU preprocessing of lattice segments enables intermediate lattice results to be returned to the requestor during streaming inference. Collectively, the proposed algorithm yields up to a 240x speedup over single core CPU decoding, and up to 40x faster decoding than the current state-of-the-art GPU decoder, while returning equivalent results. This decoder design enables deployment of production-grade ASR models on a large spectrum of systems, ranging from large data center servers to low-power edge devices.