We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Sound

Title: Weight, Block or Unit? Exploring Sparsity Tradeoffs for Speech Enhancement on Tiny Neural Accelerators

Abstract: We explore network sparsification strategies with the aim of compressing neural speech enhancement (SE) down to an optimal configuration for a new generation of low power microcontroller based neural accelerators (microNPU's). We examine three unique sparsity structures: weight pruning, block pruning and unit pruning; and discuss their benefits and drawbacks when applied to SE. We focus on the interplay between computational throughput, memory footprint and model quality. Our method supports all three structures above and jointly learns integer quantized weights along with sparsity. Additionally, we demonstrate offline magnitude based pruning of integer quantized models as a performance baseline. Although efficient speech enhancement is an active area of research, our work is the first to apply block pruning to SE and the first to address SE model compression in the context of microNPU's. Using weight pruning, we show that we are able to compress an already compact model's memory footprint by a factor of 42x from 3.7MB to 87kB while only losing 0.1 dB SDR in performance. We also show a computational speedup of 6.7x with a corresponding SDR drop of only 0.59 dB SDR using block pruning.
Comments: To appear in NeurIPS 2021 Efficient Natural Langauge and Speech Processing Workshop as oral-spotlight presentation
Subjects: Sound (cs.SD); Machine Learning (cs.LG); Audio and Speech Processing (eess.AS)
Cite as: arXiv:2111.02351 [cs.SD]
  (or arXiv:2111.02351v2 [cs.SD] for this version)

Submission history

From: Marko Stamenovic [view email]
[v1] Wed, 3 Nov 2021 17:06:36 GMT (451kb,D)
[v2] Tue, 9 Nov 2021 23:19:46 GMT (451kb,D)

Link back to: arXiv, form interface, contact.