feat(settings): make DTLN the default ML noise-suppression model
CI / Build & Quality Checks (push) Skipped
CI / Secret scan (gitleaks) (push) Skipped
CI / Playwright smoke (e2e) (push) Skipped
CI / Secret scan (gitleaks) (pull_request) Successful in 17s
CI / Build & Quality Checks (pull_request) Successful in 3m45s
CI / Trigger Desktop Build (pull_request) Skipped
CI / Docker image build & smoke test (pull_request) Skipped
CI / Playwright smoke (e2e) (pull_request) Successful in 33m3s

Browser-native stays the overall default; this only changes which model
the ML option starts on.

Measured on real speech mixed with fan, keyboard, street and babble noise
(SNR 5/15 dB), scored against the clean voice:
- DTLN: most natural voice (best PESQ/STOI of the four), strong on fans
  and typing, ~9% of one core.
- DeepFilterNet 3: removes the most noise (typing, traffic) but voices
  can sound slightly muffled, ~30% of one core.
- RNNoise: steady noise only; little help with typing.
- Speex: barely reduces noise.
- No model removes other people talking in the background.

The model list is now ordered by recommendation, with descriptions and
ratings from these results. The "Transients" column is renamed
"Typing & clicks".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5
This commit is contained in:
Lotus CI
2026-10-30 15:05:31 -04:00
co-authored by Claude Opus 5.5
parent d2c3c56b18
commit 9ab01ff6f6
4 changed files with 30 additions and 25 deletions
@@ -2037,7 +2037,7 @@ export function Calls() {
<Text size="L400">Model</Text>
<SettingTile
title="ML model"
description="Choose the machine learning model used for noise removal. Heavier models clean more aggressively at a higher CPU cost. CPU figures are the share of one processor core during a call on a mid-range desktop; Task Manager shows that divided by your number of cores, and slower computers use more."
description="Choose the machine learning model used for noise removal. DTLN is the best all-rounder; DeepFilterNet removes the most noise at a higher CPU cost. None of the models remove other people talking in the background. CPU figures are the share of one processor core during a call on a mid-range desktop; Task Manager shows that divided by your number of cores, and slower computers use more."
after={
<SettingsSelect<DenoiseModelId>
value={callDenoiseModel}
@@ -2066,7 +2066,7 @@ export function Calls() {
[
{ label: 'CPU (one core)', value: selectedDenoiseModel.cpuUsage },
{ label: 'Voice quality', value: selectedDenoiseModel.voiceQuality },
{ label: 'Transients', value: selectedDenoiseModel.transients },
{ label: 'Typing & clicks', value: selectedDenoiseModel.transients },
{ label: 'Download', value: selectedDenoiseModel.binarySize },
] as const
).map((stat) => (
@@ -2113,7 +2113,7 @@ export function Calls() {
</Box>
<Box grow="Yes">
<Text size="T200" priority="300">
Transients
Typing & clicks
</Text>
</Box>
</Box>