Audio samples from "Very Attentive Tacotron: Robust and Unbounded Length Generalization in Autoregressive Transformer-Based Text-to-Speech"

This page contains a set of audio samples in support of the paper: it is suggested that the reader listen to the samples in conjunction with reading the paper. All utterances were unseen during training.

The systems being compared are:

• Ground Truth: Human-produced audio
• Tacotron-GMMA: Tacotron with GMM-based attention.
  • Recurrent autoregressive decoder with cross-attention to the encoded phoneme sequence predicting mel spectrogram frames (L1 loss).
• NAT: Non-attentive Tacotron with unsupervised durations.
  • Recurrent autoregressive decoder with Gaussian upsampling of the encoded phoneme sequence predicting mel spectrogram frames (L1+L2 loss).
• T5-TTS: Baseline transformer-based system.
  • T5 transformer decoder with repeated multi-head cross-attentions to the encoded phoneme sequence predicting discrete audio tokens (negative log likelihood loss).
• VAT (proposed): Very Attentive Tacotron.
  • An extension of T5-TTS where the multi-head cross-attentions are informed by a single monotonic alignment position.

Contents

Examples of synthesized speech for models trained on an internal multi-speaker dataset, with the exception of Tacotron-GMMA which was trained only on the Lessac data. The utterances are taken from the collection of 885 utterances used in naturalness evaluations. Samples which deviate from the transcript are marked in red. Note that:

• T5-TTS and VAT both produce produce expressive speech of high quality on short utterances without repeated words.
• Tacotron-GMMA lacks expressivity and sometimes produces low-quality speech.

Examples of synthesized speech for models trained on the clean-460 subset of LibriTTS. The utterances are taken from the collection of 900 utterances used in naturalness evaluations. Samples which deviate from the transcript are marked in red. Note that:

• T5-TTS and VAT are similarly expressive on short utterances without repeated words.
• NAT lacks expressivity and often has highly robotic pacing.

Examples of synthesis of long utterances for models trained on the internal multi-speaker dataset. The maximum training lengths were 9.6 seconds and 192 phonemes. Samples which deviate from the transcript are marked in red. Note that:

• T5-TTS deviates from the transcript for utterances longer than those on which it was trained.
• VAT adheres to the transcript, even for extremely long utterances.
• Tacotron-GMMA adheres to the transcript but lacks expressivity.

Examples of synthesis of long utterances for models trained on LibriTTS. The maximum training lengths were 9.6 seconds and 192 phonemes. Samples which deviate from the transcript are marked in red. Note that:

• T5-TTS deviates from the transcript for utterances longer than those on which it was trained.
• VAT adheres to the transcript, even for extremely long utterances.
• NAT adheres to the transcript but lacks expressivity and often has robotic pacing.

Examples of synthesis of utterances containing repeated words for models trained on the internal multi-speaker dataset. Samples which deviate from the transcript are marked in red. The number of occurrences of the repeated word in the synthesized audio is indicated next to the play button. Note that:

• T5-TTS deviates from the transcript for utterances with repeated words.
• VAT adheres to the transcript, even for utterances with many repeated words.

Examples of synthesized speech from additional speakers for the T5-TTS and VAT models in section 1. The utterances are taken from a held-out subset of the internal multi-speaker dataset.