Audio Filter — Quality Models (V1 + V2)
Audio-quality assessment models for the Audio Filter Pipeline.
Source code: https://github.com/KaniTTS-research-team/audio_filter
What this is
Two LogisticRegression models that score how likely an audio recording is bad
quality — bad_prob, a value from 0.0 (clean) to 1.0 (bad):
- V1 — narrowband audio, sample rate <= 24 kHz. 12 DSP metrics.
- V2 — wideband audio, sample rate > 24 kHz. 34 DSP metrics.
The pipeline picks the model automatically from each file's sample rate. The
bad_prob score is turned into a good / uncertain / bad verdict by two
cut-offs per model.
Files
| File | Role |
|---|---|
model_v1/filter_model.pkl |
V1 — model + scaler + feature list |
model_v2/lr_model.pkl |
V2 — LogisticRegression model |
model_v2/scaler.pkl |
V2 — StandardScaler |
model_v2/config.json |
V2 — feature list |
Usage
These files are downloaded automatically by the
audio_filter pipeline —
you do not load them by hand. See the pipeline's README and MODELS.md for
details on how V1/V2 work and how verdicts are determined.
License
Apache 2.0.