sonique.Back to listening room

The listening room

Real-world speech.

Listen to English speech the source pipeline rejected before transcription. Compare each noisy podcast or YouTube recording with Super Sonique v4 at the same playback position.

Recordings
120
Selection
Source ASR skipped
EMA checkpoint
…
Noise tier

Selected recording

Loading recording…

Source ASR skipped
SpectrogramOriginal
Frequency over time in the selected recording
Hz0 s

Loading audio…

Source gate evidence
Source ASR status
Source skip reason
Original quality score
Gate threshold
Later good-audio MOS minimum
Crop background quality
Crop overall quality
Local speech check

How training noise is sampled

Slide view ↗

Each training example gets one augmentation. Additive noise is sampled independently by category and level.

75% additive noiseMix a noise recording into the speech.
15% room reverbSimulate sound reflecting around a room.
10% bandwidth limitingLower rates reduce high frequencies. One in six draws keeps full bandwidth.
What kind of noise?
  • SustainedA steady sound, like an engine hum.
    70%
  • MixedA background sound with louder moments, like passing traffic.
    20%
  • IntermittentShort events with quiet gaps, like hammer taps.
    10%
How loud is the noise?
  • 25%20 dB
  • 45%15 dB
  • 20%10 dB
  • 7%5 dB
  • 3%0 dB

A lower signal-to-noise ratio means louder noise relative to speech. At 0 dB, both have equal average power. Each mix varies uniformly within ±2 dB of its target.

Noise recording credits

About these clips

The source pipeline skips speech recognition below a DNSMOSPro score of 3.0. The heavy-noise tier has source scores of 2.0 or lower. On the exact excerpts, its DNSMOS P.835 background scores are at most 2.25 and overall scores are at most 2.50. The light-noise tier has both source DNSMOSPro and excerpt background scores between 2.25 and 2.80, inclusive. Lower background scores indicate more intrusive noise. Each source manifest records skipped transcription. We independently checked each excerpt for English speech.

The model output uses the eight-step Heun sampler. Both versions use the same playback gain and spectrogram scale. Press A or B while listening to switch at the same point in the recording.