Buckets:
| """Audio / control constants. They must match training: the charactr/vocos-mel-24khz mel front end.""" | |
| SR = 24000 | |
| N_FFT = 1024 | |
| HOP = 256 | |
| N_MELS = 100 | |
| FPS = SR / HOP # 93.75 mel frames per second | |
| PITCH_SCALE = 3.0 # predicted token pitch is semitones re the speaker median / 3; word_shift pitch is given in st | |
| # Utterance-level controls. Each is a SPEAKER-RELATIVE z-score: 0 = this voice's own neutral speech, +-1 = one | |
| # pooled within-speaker standard deviation (pooled_sd in control_stats.json gives the raw unit per dim). | |
| CONTROL_DIMS = [ | |
| "rate", # articulation rate, phones per second of speech (pauses excluded) | |
| "loud", # mean RMS level of speech frames, dB | |
| "f0_med", # median F0, semitones | |
| "f0_iqr", # F0 interquartile range, semitones (pitch range / liveliness) | |
| "f0_end", # F0 slope over the final 300 ms of voicing, st/s (+ rising, question-like) | |
| "pause", # share of the speech span that is silent pauses >= 100 ms | |
| "tilt", # alpha ratio: energy 1-5 kHz vs 50 Hz-1 kHz, dB (+ brighter / pressed) | |
| "voiced", # voiced share of speech frames (whisper -> low) | |
| "flat", # spectral flatness of speech frames, log (breathy / whisper -> high) | |
| "centroid", # spectral centroid of speech frames, log Hz (brightness) | |
| "cpps", # smoothed cepstral peak prominence, dB (periodicity; breathy / tense -> low) | |
| "h1h2", # H1-H2, dB (breathy -> high) | |
| ] | |
| CONTROL_RANGE = (-5.0, 5.0) # accepted request range; training labels rarely exceed +-3 | |
| DEFAULT_VOICE = "voice_1688" # v1.0.1 default (UTMOS sweep on non-test sentences); v1.0.0 and the leaderboard used voice_1361 | |
Xet Storage Details
- Size:
- 1.65 kB
- Xet hash:
- 22c44dc57ce97abdb32ef6091934faf90b1d34f2587d1799660c1914b9875230
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.