Add a delivery primer to the model card
Browse files- README.md +22 -0
- SHA256SUMS.json +2 -2
README.md
CHANGED
|
@@ -32,6 +32,28 @@ The default download is the LiteRT bundle, installed under `scyllasband/models/l
|
|
| 32 |
| Energy, tension, valence, assertiveness | 0–4 or unspecified | 2 each |
|
| 33 |
| Whisper | on/off or unspecified | off |
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
`--delivery auto` omits all coordinates. A per-axis `auto` omits just that coordinate; other unspecified fields in a partial request use neutral defaults. Omission is represented by a separate mask, not by the value 2. It is not an automatic text-to-performance planner.
|
| 36 |
|
| 37 |
The graph receives `(rating - 2) / 4` for the four continuous axes and 0/1 for whisper, plus five presence bits. The public interface accepts the original 0–4 values. Expressiveness is an auxiliary training target, not an inference axis.
|
|
|
|
| 32 |
| Energy, tension, valence, assertiveness | 0–4 or unspecified | 2 each |
|
| 33 |
| Whisper | on/off or unspecified | off |
|
| 34 |
|
| 35 |
+
With every control at 2, the default, voices speak in a calm, conversational delivery with a neutral tone. Moving a control away from 2 changes the delivery:
|
| 36 |
+
|
| 37 |
+
| Control | Toward 0 | Toward 4 |
|
| 38 |
+
| --- | --- | --- |
|
| 39 |
+
| Energy | Lower, darker voice with more pauses between phrases | Higher, brighter, more projected voice |
|
| 40 |
+
| Tension | Slightly looser and steadier | Tighter, with wider pitch swings |
|
| 41 |
+
| Valence | Flatter pitch, a little slower, longer pauses | Livelier pitch, a little quicker, shorter pauses |
|
| 42 |
+
| Assertiveness | Hesitant, with longer pauses between phrases | Faster and clipped with few pauses; the strongest control over pace |
|
| 43 |
+
|
| 44 |
+
Whisper on gives quieter, breathy speech. Combining the controls gives familiar deliveries:
|
| 45 |
+
|
| 46 |
+
| Delivery | Settings |
|
| 47 |
+
| --- | --- |
|
| 48 |
+
| Calm and soft | `energy=1.2,tension=0.8,valence=2.6` |
|
| 49 |
+
| Assertive and rapid-fire | `energy=2.6,tension=2.2,assertiveness=3.6` |
|
| 50 |
+
| Joyful | `energy=3.2,valence=3.6` |
|
| 51 |
+
| Angry | `energy=3.4,tension=3.4,valence=0.8,assertiveness=3.4` |
|
| 52 |
+
| Sad | `energy=1,tension=1.6,valence=0.8,assertiveness=1.6` |
|
| 53 |
+
| Whispered | `whisper=on` |
|
| 54 |
+
|
| 55 |
+
The [voice gallery](https://lowkeytea.github.io/scyllasband/#delivery) plays each of these for all ten voices with the exact settings used. How far a control moves the delivery varies with voice, language and text, so start near 2 and listen as you move toward the extremes.
|
| 56 |
+
|
| 57 |
`--delivery auto` omits all coordinates. A per-axis `auto` omits just that coordinate; other unspecified fields in a partial request use neutral defaults. Omission is represented by a separate mask, not by the value 2. It is not an automatic text-to-performance planner.
|
| 58 |
|
| 59 |
The graph receives `(rating - 2) / 4` for the four continuous axes and 0/1 for whisper, plus five presence bits. The public interface accepts the original 0–4 values. Expressiveness is an auxiliary training target, not an inference axis.
|
SHA256SUMS.json
CHANGED
|
@@ -4,8 +4,8 @@
|
|
| 4 |
"bytes": 11357
|
| 5 |
},
|
| 6 |
"README.md": {
|
| 7 |
-
"sha256": "
|
| 8 |
-
"bytes":
|
| 9 |
},
|
| 10 |
"litert/assets/fixed_references.json": {
|
| 11 |
"sha256": "d3c26e7e90d90fa719254ff1ff05dc4b391cd47231b7556c3c2ed5f63547925b",
|
|
|
|
| 4 |
"bytes": 11357
|
| 5 |
},
|
| 6 |
"README.md": {
|
| 7 |
+
"sha256": "b7590ee31e834406273b1bf28ea3ee2380aba97f18bf6113d39b60e745add56f",
|
| 8 |
+
"bytes": 7158
|
| 9 |
},
|
| 10 |
"litert/assets/fixed_references.json": {
|
| 11 |
"sha256": "d3c26e7e90d90fa719254ff1ff05dc4b391cd47231b7556c3c2ed5f63547925b",
|