Amazon Technologies, Inc.

Panagiota Karanasou

Thomas Renaud Drugman

Sri Vishnu Kumar Karlapati

Syed Ammar Abbas

Arnaud Vincent Pierre Yves Joly

Simon Slangen

Alexis Pierre Moinet

Edward C Bueche

Toby Ray Latin-Stoermer

Brandon Scott Durham

Fabian Moerchen

Jonathan Barnett Feinstein

Alok Verma

Fabian Triefenbach

Amina Shabbeer

Catherine Breslin

Klaus Reiter

Judith Gaspers

Alexis Pierre Jean-Baptiste Moinet

Jaime Lorenzo Trueba

Bajibabu Bollepalli

Petr Makarov

Arnaud Joly

Zack Daniel Hodari

Constanze Stypula

A speech-processing system receives input data representing text. A prosody prediction component processes the input data to determine prosody embedding data corresponding to prosody of the text. A decoder processes the prosody embedding data and phoneme encoded data derived from the input data to determine audio output data corresponding to the text and the prosody.

Synthetic speech processing related to prosody prediction

A speech-processing system receives input data representing text. One or more encoders trained to predict audio properties corresponding to the text process the text to predict those properties. A speech decoder processes phoneme embeddings as well as the predicted properties to create data representing synthesized speech.

Synthetic speech processing by representing text by phonemes exhibiting predicted volume and pitch using neural networks

Devices and techniques are generally described for learned condition text-to-speech synthesis. In some examples, first data representing a selection of a type of prosodic expressivity may be received. In some further examples, a selection of content comprising text data may be received. First audio data may be determined that includes an audio representation of the text data. The first audio data may be generated based at least in part on sampling from a first latent distribution generated using a conditional primary variational autoencoder (VAE). The sampling from the first latent distribution may be conditioned on a first learned distribution associated with the type of prosodic expressivity. In various examples, the first audio data may be sent to a first computing device.

Learned condition text-to-speech synthesis

Techniques for performing text-to-speech are described. An exemplary method includes receiving a request to generate audio from input text; generating audio from the input text by: generating a first number of vectors from phoneme embeddings representing the input text, predicting one or more spectrograms having the first number of frames using multiple scales wherein a coarser scale influences a finer scale, concatenating the first number of vectors and the predicted one or more spectrograms, generating at least one mel spectrogram from the concatenated vectors and the predicted one or more spectrograms, and converting, with a vocoder, the at least one mel spectrogram frames to audio; and outputting the generated audio according to the request.

Multi-scale spectrogram text-to-speech

A speech-processing system receives input data representing text. An input encoder processes the input data to determine first embedding data representing the text. A local attention encoder processes a subset of the first embedding data in accordance with a predicted size to determine second embedding data. An attention encoder processes the second embedding data to determine third embedding data. A decoder processes the third embedding data to determine audio data corresponding to the text.

Synthetic speech processing

Techniques are provided for training a target language model based at least in part on data associated with a reference language model. For example, language data utilized to train an English language model may be translated and provided as training data to train a German language model to recognize utterances provided in German. By utilizing the techniques herein, the efficiency of training a new language model may be improved due at least in part to replacing labor-intensive operations conventionally performed by specialized personnel with machine-generated data. Additionally, techniques discussed herein provide for reducing the time required for training a new language model by leveraging information associated with utterances of one language to train the new language model associated with a different language.

Techniques for model training for voice features

Techniques are provided for training a language recognition model. For example, a language recognition model may be maintained and associated with a reference language (e.g., English). The language recognition model may be configured to accept as input an utterance in the reference language and to identify a feature to be executed in response to receiving the utterance. New language data (e.g., other utterances) provided in a different language (e.g., German) may be obtained. This new language data may be translated to English and utilized to retrain the model to recognize reference language data as well as language data translated to the reference language. Subsequent utterances (e.g., English utterances, or German utterances translated to English) may be provided to the updated model and a feature may be identified. One or more instructions may be sent to a user device to execute a set of instructions associated with the feature.

Growing community of inventors

Cambridge, United Kingdom

Panagiota Karanasou