User testing found that DeepFilterNet "sounds choppy at the end of every word, like it cuts the mic between words slightly too soon."
At noiseReductionLevel: 60 the background is muted outright: pauses measure −61 to −64 dB, which is digital silence. Every word therefore drops straight into dead air. On top of that, DFN trims 4–9 dB off word tails in a room with echo.
Measurements
Test material: about 28 s of speech with room echo (RT60 ~0.4 s), clean and with fan and keyboard noise at SNR 15 dB, run through the same DFN build and scored against the reverberant clean speech:
Limit
PESQ clean / fan / keyboard
Noise removed
60 (old)
3.35 / 2.33 / 2.04
~62 dB
40
3.41 / 2.45 / 2.21
32–40 dB
25
3.59 / 2.72 / 2.38
25 dB
20
3.63 / 2.66 / 2.25
20 dB
25 dB still removes more than RNNoise (8–16 dB) and leaves a quiet background bed instead of silence.
Rejected alternative: a release "hangover" that blends the dry signal back in at word ends. It barely helped the tails, and keyboard clicks triggered it: keyboard noise removed fell from 61 dB to 2 dB.
## Problem
User testing found that DeepFilterNet "sounds choppy at the end of every word, like it cuts the mic between words slightly too soon."
At `noiseReductionLevel: 60` the background is muted outright: pauses measure −61 to −64 dB, which is digital silence. Every word therefore drops straight into dead air. On top of that, DFN trims 4–9 dB off word tails in a room with echo.
## Measurements
Test material: about 28 s of speech with room echo (RT60 ~0.4 s), clean and with fan and keyboard noise at SNR 15 dB, run through the same DFN build and scored against the reverberant clean speech:
| Limit | PESQ clean / fan / keyboard | Noise removed |
|---|---|---|
| 60 (old) | 3.35 / 2.33 / 2.04 | ~62 dB |
| 40 | 3.41 / 2.45 / 2.21 | 32–40 dB |
| **25** | **3.59 / 2.72 / 2.38** | 25 dB |
| 20 | 3.63 / 2.66 / 2.25 | 20 dB |
25 dB still removes more than RNNoise (8–16 dB) and leaves a quiet background bed instead of silence.
**Rejected alternative:** a release "hangover" that blends the dry signal back in at word ends. It barely helped the tails, and keyboard clicks triggered it: keyboard noise removed fell from 61 dB to 2 dB.
The in-app tester gets the same change in cinny.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5
At 60 dB the background is muted outright (measured -61 to -64 dB), so
each word drops into dead silence and listeners hear the mic cutting out
at word ends. On room-echo speech (RT60 ~0.4 s) with fan and keyboard
noise at SNR 15 dB, 25 dB scored best:
limit PESQ clean/fan/keyboard noise removed
60 3.35 / 2.33 / 2.04 ~62 dB
40 3.41 / 2.45 / 2.21 32-40 dB
25 3.59 / 2.72 / 2.38 25 dB
20 3.63 / 2.66 / 2.25 20 dB
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5
jared
scheduled this pull request to auto merge when all checks succeed 2026-10-30 18:50:25 -04:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Problem
User testing found that DeepFilterNet "sounds choppy at the end of every word, like it cuts the mic between words slightly too soon."
At
noiseReductionLevel: 60the background is muted outright: pauses measure −61 to −64 dB, which is digital silence. Every word therefore drops straight into dead air. On top of that, DFN trims 4–9 dB off word tails in a room with echo.Measurements
Test material: about 28 s of speech with room echo (RT60 ~0.4 s), clean and with fan and keyboard noise at SNR 15 dB, run through the same DFN build and scored against the reverberant clean speech:
25 dB still removes more than RNNoise (8–16 dB) and leaves a quiet background bed instead of silence.
Rejected alternative: a release "hangover" that blends the dry signal back in at word ends. It barely helped the tails, and keyboard clicks triggered it: keyboard noise removed fell from 61 dB to 2 dB.
The in-app tester gets the same change in cinny.
🤖 Generated with Claude Code
https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5