|
|
import numpy as np |
|
|
from numpy.typing import NDArray |
|
|
import json |
|
|
from phonemizer.backend.espeak.wrapper import EspeakWrapper |
|
|
from phonemizer import phonemize |
|
|
import espeakng_loader |
|
|
import onnxruntime as ort |
|
|
|
|
|
|
|
|
_BOS = "^" |
|
|
_EOS = "$" |
|
|
_PAD = "_" |
|
|
|
|
|
|
|
|
class Piper: |
|
|
def __init__( |
|
|
self, |
|
|
model_path: str, |
|
|
config_path: str, |
|
|
): |
|
|
self.setup(model_path, config_path) |
|
|
|
|
|
def setup(self, model_path, config_path, session = None): |
|
|
with open(config_path) as fp: |
|
|
self.config: dict = json.load(fp) |
|
|
self.sample_rate: int = self.config['audio']['sample_rate'] |
|
|
self.phoneme_id_map: dict = self.config['phoneme_id_map'] |
|
|
self._voices: dict = self.config.get('speaker_id_map') |
|
|
|
|
|
EspeakWrapper.set_library(espeakng_loader.get_library_path()) |
|
|
EspeakWrapper.data_path = espeakng_loader.get_data_path() |
|
|
self.sess = session or ort.InferenceSession( |
|
|
model_path, |
|
|
sess_options=ort.SessionOptions(), |
|
|
providers=['CPUExecutionProvider'] |
|
|
) |
|
|
self.sess_inputs_names = [i.name for i in self.sess.get_inputs()] |
|
|
|
|
|
@classmethod |
|
|
def from_session( |
|
|
cls, |
|
|
session: ort.InferenceSession, |
|
|
config_path: str, |
|
|
): |
|
|
instance = cls.__new__(cls) |
|
|
instance.setup(model_path='', config_path=config_path, session=session) |
|
|
return instance |
|
|
|
|
|
def create( |
|
|
self, |
|
|
text: str, |
|
|
speaker_id: str | int = None, |
|
|
is_phonemes = False, |
|
|
length_scale: int = None, |
|
|
noise_scale: int = None, |
|
|
noise_w: int = None, |
|
|
) -> tuple[NDArray[np.float32], int]: |
|
|
|
|
|
inference_cfg = self.config['inference'] |
|
|
length_scale = length_scale or inference_cfg['length_scale'] |
|
|
noise_scale = noise_scale or inference_cfg['noise_scale'] |
|
|
noise_w = noise_w or inference_cfg['noise_w'] |
|
|
|
|
|
sid = 0 |
|
|
if isinstance(speaker_id, str) and speaker_id in self._voices: |
|
|
sid = self._voices[speaker_id] |
|
|
elif isinstance(speaker_id, int): |
|
|
sid = speaker_id |
|
|
|
|
|
phonemes = text if is_phonemes else phonemize(text) |
|
|
phonemes = list(phonemes) |
|
|
phonemes.insert(0, _BOS) |
|
|
|
|
|
ids = self._phoneme_to_ids(phonemes) |
|
|
|
|
|
inputs = self._create_input(ids, length_scale, noise_w, noise_scale, sid) |
|
|
|
|
|
samples = self.sess.run(None, inputs)[0].squeeze((0,1)).squeeze() |
|
|
return samples, self.sample_rate |
|
|
|
|
|
def get_voices(self) -> dict | None: |
|
|
return self._voices |
|
|
|
|
|
def _phoneme_to_ids(self, phonemes: str) -> list[int]: |
|
|
ids = [] |
|
|
for p in phonemes: |
|
|
if p in self.phoneme_id_map: |
|
|
ids.extend(self.phoneme_id_map[p]) |
|
|
ids.extend(self.phoneme_id_map[_PAD]) |
|
|
ids.extend(self.phoneme_id_map[_EOS]) |
|
|
return ids |
|
|
|
|
|
def _create_input(self, ids: list[int], length_scale: int, noise_w: int, noise_scale: int, sid: int) -> dict: |
|
|
ids = np.expand_dims(np.array(ids, dtype=np.int64), 0) |
|
|
length = np.array([ids.shape[1]], dtype=np.int64) |
|
|
scales = np.array([noise_scale, length_scale, noise_w],dtype=np.float32) |
|
|
|
|
|
sid = np.array([sid], dtype=np.int64) if sid is not None else None |
|
|
input = { |
|
|
'input': ids, |
|
|
'input_lengths': length, |
|
|
'scales': scales, |
|
|
} |
|
|
if 'sid' in self.sess_inputs_names: |
|
|
input['sid'] = sid |
|
|
return input |
