One-pass Multiple Conformer and Foundation Speech Systems Compression and Quantization Using An All-in-one Neural Model
This work addresses the computational and storage inefficiencies in deploying multiple ASR systems, offering a practical solution for speech technology applications, though it is incremental in optimizing existing compression methods.
The paper tackles the problem of compressing and quantizing multiple automatic speech recognition (ASR) systems efficiently by proposing a one-pass neural model that constructs nested systems with varying encoder depths, widths, and quantization settings simultaneously, achieving word error rates comparable to or lower by up to 1.01% absolute than individually trained systems and up to 12.8x model size compression without significant performance loss.
We propose a novel one-pass multiple ASR systems joint compression and quantization approach using an all-in-one neural model. A single compression cycle allows multiple nested systems with varying Encoder depths, widths, and quantization precision settings to be simultaneously constructed without the need to train and store individual target systems separately. Experiments consistently demonstrate the multiple ASR systems compressed in a single all-in-one model produced a word error rate (WER) comparable to, or lower by up to 1.01\% absolute (6.98\% relative) than individually trained systems of equal complexity. A 3.4x overall system compression and training time speed-up was achieved. Maximum model size compression ratios of 12.8x and 3.93x were obtained over the baseline Switchboard-300hr Conformer and LibriSpeech-100hr fine-tuned wav2vec2.0 models, respectively, incurring no statistically significant WER increase.