app.py

import os

os.system("git clone --branch v2.3 https://github.com/DigitalPhonetics/IMS-Toucan.git toucan_codebase")
os.system("mv toucan_codebase/* .")

from run_model_downloader import download_models

download_models()

import gradio as gr
import numpy as np
import torch

from InferenceInterfaces.UtteranceCloner import UtteranceCloner
from Preprocessing.TextFrontend import ArticulatoryCombinedTextFrontend
from TrainingInterfaces.Text_to_Spectrogram.AutoAligner.Aligner import Aligner
from TrainingInterfaces.Text_to_Spectrogram.FastSpeech2.DurationCalculator import DurationCalculator


def float2pcm(sig, dtype='int16'):
    """
    https://gist.github.com/HudsonHuang/fbdf8e9af7993fe2a91620d3fb86a182
    """
    sig = np.asarray(sig)
    if sig.dtype.kind != 'f':
        raise TypeError("'sig' must be a float array")
    dtype = np.dtype(dtype)
    if dtype.kind not in 'iu':
        raise TypeError("'dtype' must be an integer type")
    i = np.iinfo(dtype)
    abs_max = 2 ** (i.bits - 1)
    offset = i.min + abs_max
    return (sig * abs_max + offset).clip(i.min, i.max).astype(dtype)


class TTS_Interface:

    def __init__(self):
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.utterance_cloner = UtteranceCloner(model_id="Meta", device=self.device)
        self.utterance_cloner.tts.set_language("de")
        self.acoustic_model = Aligner()
        self.acoustic_model.load_state_dict(torch.load("Models/Aligner/aligner.pt", map_location='cpu')["asr_model"])
        self.acoustic_model = self.acoustic_model.to(self.device)
        self.dc = DurationCalculator(reduction_factor=1)
        self.text = "Quellen hattest du ihm, hattest dem Flüchtigen, kühle Schatten geschenkt, und die Gestade sahen, all ihm nach, und es bebte, aus den Wellen ihr lieblich Bild."
        reference_audio = "reference_audios/2.wav"
        self.duration, self.pitch, self.energy, _, _ = self.utterance_cloner.extract_prosody(self.text, reference_audio, lang="de", on_line_fine_tune=True)
        self.utterance_cloner.tts.text2phone.use_word_boundaries = False
        self.phones = self.utterance_cloner.tts.text2phone.get_phone_string(self.text)

        #######
        self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_male.wav")
        self.current_voice = "male"
        self.cloned_speech_male = self.utterance_cloner.tts(self.phones,
                                                            view=False,
                                                            durations=self.duration,
                                                            pitch=self.pitch,
                                                            energy=self.energy,
                                                            phones=True).cpu().numpy()
        self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_female.wav")
        self.current_voice = "female"
        self.cloned_speech_female = self.utterance_cloner.tts(self.phones,
                                                              view=False,
                                                              durations=self.duration,
                                                              pitch=self.pitch,
                                                              energy=self.energy,
                                                              phones=True).cpu().numpy()

        #######
        self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_male.wav")
        self.current_voice = "male"
        self.reg_speech_male = self.utterance_cloner.tts(
            "Quellen hattest du ihm, hattest dem Flüchtigen kühle Schatten geschenkt, und die Gestade sahen all ihm nach, und es bebte aus den Wellen ihr lieblich Bild.",
            view=False).cpu().numpy()
        self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_female.wav")
        self.current_voice = "female"
        self.reg_speech_female = self.utterance_cloner.tts(
            "Quellen hattest du ihm, hattest dem Flüchtigen kühle Schatten geschenkt, und die Gestade sahen all ihm nach, und es bebte aus den Wellen ihr lieblich Bild.",
            view=False).cpu().numpy()

    def read(self, _, speaker, lengthening, pause_dur, pitch_up):

        if speaker == "Female Voice" and self.current_voice != "female":
            self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_female.wav")
            self.current_voice = "female"
        elif speaker == "Male Voice" and self.current_voice != "male":
            self.utterance_cloner.tts.set_utterance_embedding("reference_audios/german_male.wav")
            self.current_voice = "male"

        duration = self.duration.clone()
        # lengthening
        lenghtening_candidates = [
            # ('f', 27),
            # ('l', 28),
            ('ʏ', 29),
            ('ç', 30),
            # ('t', 31),
            ('ɪ', 32),
            # ('ɡ', 33),
            ('ə', 34),
            ('n', 35),

            # ('z', 66),
            ('ɑ', 67),
            # ('ə', 68),
            ('n', 69),

            # ('b', 84),
            ('e', 85),
            # ('p', 86),
            # ('t', 87),
            ('ə', 88)
            ]

        for lenghtening_candidate in lenghtening_candidates:
            duration[lenghtening_candidate[1]] = duration[lenghtening_candidate[1]] + lengthening

        # pauses
        pause_candidates = [('~', 36),
                            ('~', 70),
                            ('~', 89)]

        for pause_candidate in pause_candidates:
            duration[pause_candidate[1]] = duration[pause_candidate[1]] + pause_dur

        pitch = self.pitch.clone()
        # pitch raise

        pitch_candidates = [
            # ('k', 37),
            ('y', 38),
            ('l', 39),
            ('ə', 40),
            ('ʃ', 41),
            ('a', 42),
            ('t', 43),
            # ('ə', 44),
            # ('n', 45),

            ('a', 71),
            ('l', 72),

            ('v', 96),
            ('ɛ', 97),
            ('l', 98),
            # ('ə', 99),
            # ('n', 100)
            ]

        for pitch_candidate in pitch_candidates:
            pitch[pitch_candidate[1]] = pitch[pitch_candidate[1]] + pitch_up

        fixme = [('f', 27),
                 ('l', 28),
                 ('ʏ', 29),
                 ('ç', 30),
                 ('t', 31),
                 ('ɪ', 32),
                 ('ɡ', 33),
                 ('ə', 34),
                 ('n', 35)
                 ]
        for pitch_candidate in fixme:
            pitch[pitch_candidate[1]] = pitch[pitch_candidate[1]] - abs(pitch_up)

        manipulated_speech = self.utterance_cloner.tts(self.phones,
                                                       view=False,
                                                       durations=duration,
                                                       pitch=pitch,
                                                       energy=self.energy,
                                                       input_is_phones=True).cpu()

        if self.current_voice == "female":
            cloned_speech = self.cloned_speech_female
            reg_speech = self.reg_speech_female
        else:
            cloned_speech = self.cloned_speech_male
            reg_speech = self.reg_speech_male

        return (48000, float2pcm(reg_speech)), (48000, float2pcm(cloned_speech)), (48000, float2pcm(manipulated_speech.numpy()))


poem_model = TTS_Interface()
article = "<p style='text-align: left'>This is still a work in progress, models will be exchanged for better ones as soon as they are done. More diverse training data can help with more exact cloning and more controllability. For example we are still trying to incorporate more singing data. </p><p style='text-align: center'><a href='https://github.com/DigitalPhonetics/IMS-Toucan' target='_blank'>Click here to learn more about the IMS Toucan Speech Synthesis Toolkit</a></p>"

iface = gr.Interface(fn=poem_model.read,
                     inputs=[gr.inputs.Dropdown([
                         "Quellen hattest du ihm, hattest dem Flüchtigen // kühle Schatten geschenkt, und die Gestade sahn // all ihm nach, und es bebte // aus den Wellen ihr lieblich Bild."],
                         type="value",
                         default="Quellen hattest du ihm, hattest dem Flüchtigen // kühle Schatten geschenkt, und die Gestade sahn // all ihm nach, und es bebte // aus den Wellen ihr lieblich Bild.",
                         label="Poem Transcript"),
                         gr.inputs.Dropdown(["Female Voice", "Male Voice"],
                                            type="value",
                                            default="Female Voice",
                                            label="Select a Speaker"),
                         gr.inputs.Slider(minimum=0, maximum=4, step=1, default=2, label="Lengthening on verse end"),
                         gr.inputs.Slider(minimum=0, maximum=20, step=1, default=8, label="Length of Pause after verse end"),
                         gr.inputs.Slider(minimum=-0.4, maximum=0.4, step=0.01, default=0.2, label="Raise Pitch on new verse")
                         ],
                     outputs=[gr.outputs.Audio(type="numpy", label="Poem read with prose reading"),
                              gr.outputs.Audio(type="numpy", label="Poem cloned from a reference"),
                              gr.outputs.Audio(type="numpy", label="Poem after human-in-the-loop adjustments")],
                     layout="vertical",
                     title="PoeticTTS - Customizing Poetry",
                     thumbnail="Utility/toucan.png",
                     theme="default",
                     allow_flagging="never",
                     allow_screenshot=False,
                     description="Customize how a poem is read by a text-to-speech system with intuitive high-level controls. You can control phrasing markers to go from prose style syntactic phrasing to verse aware poetry style phrasing with the sliders below.",
                     article=article)
iface.launch(enable_queue=True)