Rileypiper commited on
Commit
04fce0e
·
verified ·
1 Parent(s): 71d3208

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +21 -31
README.md CHANGED
@@ -1,4 +1,14 @@
1
- # Voice Target Slider
 
 
 
 
 
 
 
 
 
 
2
 
3
  Upload a reference voice to see where it lands on five acoustic dimensions, then record your own to compare. Designed for gender-affirming voice therapy.
4
 
@@ -14,35 +24,15 @@ Upload a reference voice to see where it lands on five acoustic dimensions, then
14
 
15
  ## Accuracy Features
16
 
17
- ### Pitch Extraction
18
- - **Cross-correlation with `very_accurate=True`** — doubles the analysis window, reducing octave errors
19
-
20
- ### Vocal Weight (H1*-H2*)
21
- - **Iseli-Alwan (2004) formant correction** — raw H1-H2 is biased when F0 is near F1. H1* subtracts the vocal tract transfer function to isolate true source spectral tilt
22
- - **4096-point FFT** at 22050 Hz (3.9 Hz resolution) — halves quantization error in harmonic peak picking
23
-
24
- ### Resonance
25
- - **Adaptive max_formant** — adjusts based on detected F0 range per Praat manual
26
- - **Formant bandwidth vowel gating** — filters out nasals/approximants (B1 > 400 Hz)
27
- - **F3 tracking + formant dispersion** — (F3-F1)/2 is more robust to vowel identity than F2-F1 alone (Fitch 1997)
28
- - **VTL estimation** — vocal tract length from speed of sound / (4 x delta_f)
29
- - **Composite z-score** — F2-F1 and dispersion both z-scored against norms
30
-
31
- ### Prosody
32
- - **Pitch velocity** (st/s) and **directional change rate** (/s) for temporal dynamics
33
- - **Weighted composite** — 60% SD + 25% velocity + 15% direction changes
34
-
35
- ### Voice Quality
36
- - **Real CPP** from power cepstrum (Hillenbrand 1994) replacing HNR proxy
37
  - **Jitter and shimmer** from Praat point process
38
- - **HNR** retained as supplementary metric
39
-
40
- ### Distribution Anchors
41
- From published norms: Titze 1994, Hillenbrand 1995, Keating & Esposito 2006, Fitch 1997, Pépiot 2014
42
-
43
- ## Deploy
44
 
45
- ```bash
46
- pip install -r requirements.txt
47
- python app.py
48
- ```
 
1
+ ---
2
+ title: Voice Target Slider
3
+ emoji: 🎙️
4
+ colorFrom: indigo
5
+ colorTo: pink
6
+ sdk: gradio
7
+ sdk_version: "5.12.0"
8
+ python_version: "3.11"
9
+ app_file: app.py
10
+ pinned: false
11
+ ---
12
 
13
  Upload a reference voice to see where it lands on five acoustic dimensions, then record your own to compare. Designed for gender-affirming voice therapy.
14
 
 
24
 
25
  ## Accuracy Features
26
 
27
+ - **Cross-correlation pitch with `very_accurate=True`** — reduces octave errors
28
+ - **H1*-H2* Iseli-Alwan (2004) formant correction** — removes F1 proximity bias in spectral tilt
29
+ - **4096-point FFT** at 22050 Hz for 3.9 Hz frequency resolution
30
+ - **Adaptive max_formant** based on detected F0 range
31
+ - **Formant bandwidth vowel gating** (B1 < 400 Hz) to exclude nasals
32
+ - **F3 + formant dispersion + VTL estimation** for vowel-robust resonance
33
+ - **Real CPP** from power cepstrum (Hillenbrand 1994)
34
+ - **Prosody composite**: F0 SD (60%) + pitch velocity (25%) + direction changes (15%)
 
 
 
 
 
 
 
 
 
 
 
 
35
  - **Jitter and shimmer** from Praat point process
36
+ - **Weighted 7-component confidence** including duration and jitter
 
 
 
 
 
37
 
38
+ Normative anchors from Titze 1994, Hillenbrand 1995, Keating & Esposito 2006, Fitch 1997, Pépiot 2014.