Parallel tacotron non-autoregressive and controllable TTS

Patent metadata

Field	Value
Publication number	US-12488780-B2
Application number	US-202418421116-A
Country	US
Kind code	B2
Filing date	Jan 24, 2024
Priority date	Oct 21, 2020
Publication date	Dec 2, 2025
Grant date	Dec 2, 2025

Abstract

A method for training a non-autoregressive TTS model includes receiving training data that includes a reference audio signal and a corresponding input text sequence. The method also includes encoding the reference audio signal into a variational embedding that disentangles the style/prosody information from the reference audio signal and encoding the input text sequence into an encoded text sequence. The method also includes predicting a phoneme duration for each phoneme in the input text sequence and determining a phoneme duration loss based on the predicted phoneme durations and a reference phoneme duration. The method also includes generating one or more predicted mel-frequency spectrogram sequences for the input text sequence and determining a final spectrogram loss based on the predicted mel-frequency spectrogram sequences and a reference mel-frequency spectrogram sequence. The method also includes training the TTS model based on the final spectrogram loss and the corresponding phoneme duration loss.

First claim

What is claimed is:

1 . A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving a text utterance comprising a sequence of phonemes; obtaining a plurality of different types of positional embeddings representing phoneme position information for each phoneme in the sequence of phonemes, the plurality of different types of positional embeddings comprising a first sinusoidal embedding of a frame position within a respective phoneme, a second sinusoidal embedding of phoneme duration, and a fractional progression of a frame within a respective phoneme; obtaining a variational embedding that specifies an intended prosody/style for synthesizing the text utterance into speech; encoding, using a text encoder of a non-autoregressive text-to-speech (TTS) model, the text utterance into an encoded text sequence; for each corresponding phoneme in the sequence of phonemes, predicting, using a duration decoder of the non-autoregressive TTS model, a phoneme duration of the corresponding phoneme based on the encoded text sequence and the obtained variational embedding; upsampling, using the predicted phoneme durations, an output of the duration decoder into a number of frames; generating a combination of the plurality of different types of positional embeddings and the upsampled output of the duration decoder; generating, as output from a spectrogram decoder of the non-autoregressive TTS model, a predicted mel-frequency spectrogram sequence for the text utterance based on the combination of the plurality of different types of positional embeddings and the upsampled output of the duration decoder, the predicted mel-frequency spectrogram sequence having the intended prosody/style specified by the variational embedding; and converting, using a synthesizer, the predicted mel-frequency spectrogram sequence for the text utterance into a time-domain audio waveform indicative of synthesized speech.

2 . The computer-implemented method of claim 1 , wherein the operations further comprise providing, for audible output from an audio output device, the time-domain audio waveform indicative of the synthesized speech.

3 . The computer-implemented method of claim 1 , wherein the spectrogram decoder of the non-autoregressive TTS model comprises a stack of self-attention blocks.

4 . The computer-implemented method of claim 1 , wherein the text utterance comprises words each having one or more phonemes from the sequence of phonemes, silences at all word boundaries, and punctuation marks.

5 . The computer-implemented method of claim 1 , wherein the operations further comprise:
concatenating the encoded text sequence, the variational embedding, and a reference speaker embedding representing an identity of a reference speaker; and generating the output of the duration decoder based on the duration decoder receiving, as input, the concatenation of the encoded text sequence, the variational embedding, and the reference speaker embedding.

6 . The computer-implemented method of claim 1 , wherein encoding the text utterance into the encoded text sequence comprises:
receiving, from a phoneme look-up table, a respective embedding of each phoneme in the sequence of phonemes; for each phoneme in the sequence of phonemes, processing, using an encoder pre-net neural network of the text encoder, the respective embedding to generate a respective transformed embedding of the phoneme; processing, using a bank of convolutional blocks, the respective transformed embeddings to generate convolution outputs; and processing, using a stack of self-attention blocks, the convolution outputs to generate the encoded text sequence.

7 . The computer-implemented method of claim 1 , wherein predicting the phoneme duration for the corresponding phoneme comprises:
predicting a probability of non-zero duration for the corresponding phoneme; predicting a continuous phoneme duration for the corresponding phoneme; determining the probability of non-zero duration predicted for the corresponding phoneme is less than a threshold value; and based on determining the probability of non-zero duration predicted for the corresponding phoneme is less than the threshold value, zeroing out the continuous phoneme duration predicted for the corresponding phoneme.

8 . The computer-implemented method of claim 1 , wherein predicting the phoneme duration for the corresponding phoneme comprises:
predicting a probability of non-zero duration for the corresponding phoneme; predicting a continuous phoneme duration for the corresponding phoneme; determining the probability of non-zero duration predicted for the corresponding phoneme is not less than a threshold value; and based on determining the probability of non-zero duration predicted for the corresponding phoneme is not less than the threshold value, setting a value of the predicted phoneme duration for the corresponding phoneme equal to a value of the continuous phoneme duration predicted for the corresponding phoneme.

9 . The computer-implemented method of claim 1 , wherein the duration decoder comprises a stack of self-attention blocks followed by two independent projections.

10 . A system comprising:
data processing hardware; and memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed by the data processing hardware cause the data processing hardware to perform operations comprising:
receiving a text utterance comprising a sequence of phonemes;
obtaining a plurality of different types of positional embeddings representing phoneme position information for each phoneme in the sequence of phonemes, the plurality of different types of positional embeddings comprising a first sinusoidal embedding of a frame position within a respective phoneme, a second sinusoidal embedding of phoneme duration, and a fractional progression of a frame within a respective phoneme;
obtaining a variational embedding that specifies an intended prosody/style for synthesizing the text utterance into speech;
encoding, using a text encoder of a non-autoregressive text-to-speech (TTS) model, the text utterance into an encoded text sequence;
for each corresponding phoneme in the sequence of phonemes, predicting, using a duration decoder of the non-autoregressive TTS model, a phoneme duration of the corresponding phoneme based on the encoded text sequence and the obtained variational embedding;
upsampling, using the predicted phoneme durations, an output of the duration decoder into a number of frames;
generating a combination of the plurality of different types of positional embeddings and the upsampled output of the duration decoder;
generating, as output from a spectrogram decoder of the non-autoregressive TTS model, a predicted mel-frequency spectrogram sequence for the text utterance based on the combination of the plurality of different types of positional embeddings and the upsampled output of the duration decoder, the predicted mel-frequency spectrogram sequence having the intended prosody/style specified by the variational embedding; and
converting, using a synthesizer, the predicted mel-frequency spectrogram sequence for the text utterance into a time-domain audio waveform indicative of synthesized speech.

11 . The system of claim 10 , wherein the operations further comprise providing, for audible output from an audio output device, the time-domain audio waveform indicative of the synthesized speech.

12 . The system of claim 10 , wherein the spectrogram decoder of the non-autoregressive TTS model comprises a s

Assignees

• Google Llc

Inventors

• Elias Isaac
• Shen Jonathan
• Zhang Yu
• Jia Ye
• Weiss Ron J
• Wu Yonghui
• Chun Byungha

Classifications

• G10L13/047 Primary

  Architecture of speech synthesisers · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title