Open-source State-of-the-art TTS model which runs on a CPU 😻
Python
import pytest
import torch
from kittenml.kittentts2.logits import build_logits_processors
from kittenml.kittentts2.tokens import TokenMap
TOKEN_MAP = TokenMap(audio_id_base=100, num_audio_tokens=6561, speech_start_id=1,
speech_end_id=2, text_start_id=3, text_end_id=4, start_id=5,
stop_id=6)
SILENCE = TOKEN_MAP.audio_id_base + 4299
def test_repetition_penalty_covers_silence_but_not_terminators():
"""Silence runs are the main checkpoint's stuck state; only the stop tokens are exempt."""
processors = build_logits_processors(TOKEN_MAP, temperature=1.0, top_k=0, top_p=1.0,
min_p=0.0, repetition_penalty=1.1,
repetition_window=50, token_run_penalty=0,
prompt_length=1)
history = torch.tensor([[TOKEN_MAP.start_id, SILENCE, TOKEN_MAP.speech_end_id,
TOKEN_MAP.stop_id, 150]])
scores = torch.full((1, TOKEN_MAP.audio_id_end), 2.0)
out = processors(history, scores.clone())
assert out[0, SILENCE].item() == pytest.approx(2.0 / 1.1)
assert out[0, 150].item() == pytest.approx(2.0 / 1.1)
assert out[0, TOKEN_MAP.speech_end_id] == 2.0
assert out[0, TOKEN_MAP.stop_id] == 2.0