-
Notifications
You must be signed in to change notification settings - Fork 117
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
PiperOrigin-RevId: 681461833
- Loading branch information
Sound Separation Team
committed
Oct 2, 2024
1 parent
e041a64
commit 678e737
Showing
14 changed files
with
143 additions
and
0 deletions.
There are no files selected for viewing
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,143 @@ | ||
| <html> | ||
| <head> | ||
| <title>Towards sub-millisecond latency real-time speech enhancement models on hearables (Online Supplement)</title> | ||
| <link | ||
| href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" | ||
| rel="stylesheet" | ||
| /> | ||
| <meta charset="utf-8" /> | ||
| <meta name="viewport" content="width=device-width, initial-scale=1" /> | ||
| <style> | ||
| audio { | ||
| /* width: 250px; */ | ||
| width: 20vw; | ||
| min-width: 100px; | ||
| max-width: 250px; | ||
| } | ||
| img { | ||
| width: 100%; | ||
| /* min-height: 300px; */ | ||
| min-width: 300px; | ||
| max-width: 800px; | ||
| display: block; | ||
| margin-left: auto; | ||
| margin-right: auto; | ||
| } | ||
| td { | ||
| text-align: center; | ||
| vertical-align: middle; | ||
| } | ||
| .container { | ||
| overflow-x: auto; | ||
| } | ||
| /* Solid border */ | ||
| hr.solid { | ||
| border-top: 2px solid #bbb; | ||
| } | ||
| </style> | ||
| </head> | ||
| <body> | ||
| <div class="container pt-5 mt-5 shadow p-5 mb-5 bg-white rounded"> | ||
| <div class="text-center"> | ||
| <h1 style="text-align:center; font-size: 40px;">Towards sub-millisecond latency real-time speech enhancement models on hearables</h1> | ||
| <!-- <h2 style="text-align:center; font-size: 30px;">Online Supplement</h2> --> | ||
| <p class="lead fw-bold"> | ||
| |<a | ||
| href="https://arxiv.org/abs/2409.18239" | ||
| class="btn border-white bg-white fw-bold" | ||
| >Paper</a | ||
| >| | ||
| <br> | ||
| <p class="fst-italic mb-0"> | ||
| Artem Dementyev, Chandan K. A. Reddy, Scott Wisdom, Navin Chatlani, John R. Hershey, Richard F. Lyon | ||
| </p> | ||
| <br> | ||
| <p> | ||
| <br> | ||
|
|
||
| </p> | ||
|
|
||
| </div> | ||
| <p> | ||
| <h1>Overview</h1> | ||
| <p> | ||
| Low latency models are critical for real-time speech enhancement applications, such as hearing aids and hearables. However, the sub-millisecond latency space for resource-constrained hearables remains underexplored. We demonstrate speech enhancement using a computationally efficient minimum-phase FIR filter, enabling sample-by-sample processing to achieve mean algorithmic latency of 0.32 ms to 1.25 ms. With a single microphone, we observe a mean SI-SDRi of 4.1 dB. The approach shows generalization with a DNSMOS increase of 0.2 on unseen audio recordings. We use a lightweight LSTM-based model of 644k parameters to generate FIR taps. We benchmark that our system can run on low-power DSP with 388 MIPS and mean end-to-end latency of 3.35 ms. We provide a comparison with baseline low-latency spectral masking techniques. We hope this work will enable a better understanding of latency and can be used to improve the comfort and usability of hearables. | ||
| </p> | ||
| <h2>Deep FIR achitecture</h2> | ||
| <p> | ||
| Below is Deep FIR signal processing diagram, divided into synthesis and analysis. A new FIR filter is estimated every hop. | ||
| </p> | ||
| <img src="figs/graph_latency_v3.png" style="width: 50%;"> | ||
| </div> | ||
|
|
||
|
|
||
|
|
||
| <div class="container pt-5 mt-5 shadow p-5 mb-5 bg-white rounded"> | ||
| <div class="container pt-3"> | ||
|
|
||
|
|
||
|
|
||
| <h2>Results with 0.125 ms synthesis window Deep FIR model</h2> | ||
| <p> | ||
| The table below provides audio samples evaluated on the CHiME-2 WSJ0 test set. The minimum phase Deep FIR filter has a mean algorithmic latency of 0.38 ms, which can potentially allow for 1.6 ms end-to-end latency on hardware. | ||
| Notice that the proposed Deep FIR approach achieves comparable audio quality and more efficient computation in the very low latency range compared to the long short time window (LSTW) approach. | ||
| </p> | ||
| <table align="center"> | ||
| <tr> | ||
| <td></td> | ||
| <td colspan="1">Example -3dB SNR</td> | ||
| <td colspan="1">Example +3 dB SNR</td> | ||
| <td colspan="1">Example +9 dB SNR</td> | ||
| </tr> | ||
| <!-- '#4f2a58', '#733661', '#954466', '#b45668', '#cf6d68', '#e48767', '#f2a568', '#fac46e'--> | ||
| <tr bgcolor="#0000ff"> | ||
| <td bgcolor="#0000ff"><font color="#ffffff">Noisy Mixture</font></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/m3db_mix_440c020b.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/3db_mix_442c020q.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/9db_mix_446c020t.wav" type="audio/wav"> </audio></td> | ||
| <tr bgcolor="#4f2a58"> | ||
| <td bgcolor="#4f2a58"><font color="#ffffff">LSTW (0.5 ms alg. latency)</font></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/m3db_LSTW_440c020b.wav"" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/3db_LSTW_442c020q.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/9db_LSTW_446c020t.wav" type="audio/wav"> </audio></td> | ||
| </tr> | ||
| <tr bgcolor="#cf6d68"> | ||
| <td bgcolor="#cf6d68"><font color="#ffffff">Proposed: Deep FIR Minimum phase (0.38 ms alg. latency)<br/></font></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/m3db_MP_440c020b.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/3db_MP_442c020q.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/9db_MP_446c020t.wav" type="audio/wav"> </audio></td> | ||
| </tr> | ||
| <tr bgcolor="#fac46e"> | ||
| <td bgcolor="#fac46e"><font color="#222222">Ground-Truth Reference</font></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/m3db_original_440c020b.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/3db_original_442c020q.wav" type="audio/wav"> </audio></td> | ||
| <td colspan="1"><audio controls style="width: 175px;"> <source src="audio/9db_original_446c020t.wav" type="audio/wav"> </audio></td> | ||
| </tr> | ||
| <tr> | ||
| <td colspan="1"></td> | ||
| <td colspan="3"><hr class="solid"></td> | ||
| <!-- <td colspan="2"></td> --> | ||
| </tr> | ||
| </table> | ||
|
|
||
|
|
||
|
|
||
| </div> | ||
| </div> | ||
|
|
||
|
|
||
|
|
||
| <script> | ||
| let players = Array.prototype.slice.call(document.querySelectorAll('audio')); | ||
| players.forEach(player => { | ||
| player.addEventListener('play', event => { | ||
| players.forEach(p => { | ||
| if (player != p) {p.pause();} | ||
| }); | ||
| player.currentTime = 0.1; | ||
| }); | ||
| }); | ||
| </script> | ||
| </body> | ||
| </html> |