scyllasband / pytorch /release_config.json
spybyscript's picture
Scylla's Band v4-20261106: speech and singing from one model (a delivery mode on the flow, a BigVGAN singing decoder); LiteRT, ONNX, WebGPU, Core ML, Core AI and PyTorch
b0fe862 verified
Raw History Blame Contribute Delete
8.91 kB
{
"schema": "scyllasband_pytorch_release_v4",
"files": {
"bigvgan_adapter.pt": "77ec3dd196a7b498320ad559e340240d1c6ff6e6de96ef04690453a083134a49",
"bigvgan_v2.pt": "bf8c161493a67ee1b399b4ebac9b3119d2d940c676a19633717ee9ee31eddd64",
"condition_index.json": "2fa8376efc09e5992f008a4e72bf14150eb1c7d340eeab0691a07e04ba6a8bb2",
"duration.pt": "84e477289a4c19dc030308d25bd9784455cef916a707030bf0dca3bacde55ab8",
"fixed_references.json": "523eb475976c4749a8b6bab1ecce982f147bc2b576b3b35544df8173fa35c57a",
"g2p.pt": "1cf9c6cb570558073e55db374f2ddec2cfb8febc3ea0ecc555620c9c7c5d99f1",
"phone_vocab.json": "83e6457a69948edbd9a94b404c22fd408c263eb0d0b5acaafa4926e994587721",
"pitch_predictor.pt": "932faa2b5374a4233253b020c0e399ff20bd853fa3b52bb8abc0f49777a24d9f",
"vector_estimator.pt": "387d51db4978b2cd4fdf87f5cb3a38dbf67c75ffb08be3ab1059919cb47ac10d",
"vocoder.pt": "4bc62614f579e245a00d06ae9cc259ca177c35b8713b042af7bb09709f6ba9cb"
},
"sampler": {
"default": "euler",
"steps": 4
},
"latent": {
"dim": 24,
"hop_length": 512,
"sample_rate": 24000
},
"controls": {
"delivery": {
"defaults": {
"ariadne": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"felix": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"gwen": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"ink": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"kincade": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"max": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"orpheus": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"poppy": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"priya": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"rex": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"scylla": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"stone": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"thistle": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
},
"tuesday": {
"intensity": 0.0,
"pace": 0.0,
"pitch": 0.0
}
},
"graph_slots": 5,
"min_spoken_phones": 4,
"ranges": {
"intensity": [
-2.0,
2.0
],
"pace": [
-2.0,
2.0
],
"pitch": [
-2.0,
2.0
]
},
"release_id": "v4-20261106",
"rule": "value = clip(z, -z_limit, z_limit) / z_scale, present; both slots absent on a target with fewer than min_spoken_phones spoken phones; slots 2-4 always absent",
"schema": "measured_controls_v1",
"slots": [
"intensity",
"pace"
],
"units": {
"intensity": "sd",
"pace": "sd",
"pitch": "semitones"
},
"voice_locales": {
"ariadne": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"felix": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"gwen": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"ink": [
"de",
"en_gb",
"es",
"fr",
"it",
"vi"
],
"kincade": [
"en_us",
"es",
"de",
"fr",
"it",
"vi"
],
"max": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"orpheus": [
"de",
"en_gb",
"es",
"fr",
"it",
"vi"
],
"poppy": [
"en_us",
"es",
"de",
"fr",
"it",
"vi"
],
"priya": [
"en_us",
"es",
"de",
"fr",
"it",
"vi"
],
"rex": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"scylla": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"stone": [
"de",
"en_us",
"es",
"fr",
"it",
"vi"
],
"thistle": [
"en_us",
"es",
"de",
"fr",
"it",
"vi"
],
"tuesday": [
"de",
"en_gb",
"es",
"fr",
"it",
"vi"
]
},
"z_limit": 2.5,
"z_scale": 5.0
},
"pace_pauses": {
"per_sd": -0.13,
"phone": "<sil>",
"rule": "before rounding, each <sil> position's predicted frames are multiplied by exp(per_sd * pace)"
},
"pitch": {
"center": 10.0,
"channels": 2,
"context_encoder": "pitch_context_encoder",
"edges": {
"ariadne": [
7.69,
27.31
],
"felix": [
-3.02,
18.91
],
"gwen": [
4.84,
25.68
],
"ink": [
3.83,
24.19
],
"kincade": [
-6.17,
18.61
],
"max": [
-9.9,
17.2
],
"orpheus": [
-4.35,
18.42
],
"poppy": [
6.97,
25.88
],
"priya": [
6.0,
25.96
],
"rex": [
-3.75,
17.13
],
"scylla": [
8.47,
28.12
],
"stone": [
-7.92,
17.8
],
"thistle": [
6.5,
26.67
],
"tuesday": [
6.13,
24.86
]
},
"knee": 2.0,
"melody_center": 16.65,
"melody_scale": 24.0,
"rule": "contour [2, T]: st = c0 * scale + center + shift; st soft-limited inside the voice's edges (knee = min(knee, (high - low) / 4): unchanged until knee semitones from an edge, then edge -/+ knee * tanh); p = clip((c1 + 1) / 2, 0, 1); voiced = p > voiced_threshold; melody [T, 3] = ((st - melody_center) / melody_scale * voiced, p, voiced)",
"sampler": {
"choices": [
"euler",
"heun"
],
"default": "euler"
},
"scale": 6.0,
"steps": {
"default": 4,
"max": 64,
"min": 1
},
"voiced_threshold": 0.5
},
"delivery_modes": {
"default": "spoken",
"enabled": true,
"ids": {
"ballad": 3,
"rap": 2,
"spoken": 0,
"sung": 1
},
"input": "delivery_mode_id",
"singing": {
"decoder": "bigvgan",
"guidance": 1.0,
"melody": {
"layout": "note_phase_presence",
"midi_center": 60,
"midi_scale": 24,
"rule": "per frame: ((midi - midi_center) / midi_scale, phase 0..1 inside the note, 1 where a note sounds); zeros on rests"
},
"modes": [
"sung",
"ballad"
],
"pitch_predictor": "bypassed",
"pitch_soft_limit": "bypassed",
"rule": "sung and ballad targets take their melody from the score (expanded_melody in the note_phase_presence layout), skip the pitch predictor and the pitch soft limit, and render with the singing sampler and steps; spoken targets are unchanged from v3",
"sampler": "euler",
"steps": 4,
"voice_midi_center": {
"ariadne": 59.0,
"felix": 50.0,
"gwen": 56.0,
"ink": 57.0,
"kincade": 49.0,
"max": 49.0,
"orpheus": 50.0,
"poppy": 58.0,
"priya": 58.0,
"rex": 49.0,
"scylla": 59.0,
"stone": 49.0,
"thistle": 59.0,
"tuesday": 58.0
}
}
},
"decoders": {
"bigvgan": {
"component": "vocoder_bigvgan",
"license": "BigVGAN v2 (NVIDIA, MIT) fine-tuned",
"output_samples": "latent_frames * 512",
"required": false,
"samples_per_frame": 512
},
"vocos": {
"component": "vocoder",
"output_samples": "latent_frames * 512 - 256",
"samples_per_frame": 512
}
}
}
}