Web-Based Speech Synthesis Editor

Martin Grůber, Jakub Vít, Jindřich MatoušekInterspeech 2019

TL;DR

This paper introduces a web-based editor for speech synthesis that allows users to fine-tune synthesized speech by modifying the synthesis process itself or by post-processing the audio. This tool aims to improve the quality of synthetic speech for applications where pre-recorded phrases are acceptable, offering a more flexible and user-friendly approach than traditional methods.

What Problem Does This Paper Address?

Speech synthesis (Text-to-Speech, TTS) is increasingly used in various applications, from public announcements to e-learning. For many of these use cases, the synthesized speech doesn't need to be generated in real-time; instead, pre-prepared phrases are sufficient. The challenge lies in ensuring these pre-prepared phrases are of the highest possible quality, free from glitches, with natural prosody, and potentially fitting specific time constraints. Traditional TTS systems can produce artifacts or prosodic issues, and manually correcting these can be time-consuming and require specialized expertise. This paper addresses the need for a tool that allows for interactive refinement of synthesized speech, bridging the gap between automated synthesis and manual audio editing.

Methods

The core of this work is a web-based Graphical User Interface (GUI) that acts as a frontend to a backend TTS system called ARTIC. The system supports two primary speech synthesis methods:

  • Unit Selection: This mature technique concatenates small speech units (like phonemes or diphones) from a large corpus to form utterances. The system allows users to influence the selection of these units, potentially choosing sequences that are not objectively optimal but subjectively sound better, or to address issues like glitches. Users can also modify criteria that drive unit selection, such as speed and pitch, to influence the choice of units.
  • Statistical Parametric Synthesis (SPS): This method leverages Deep Neural Networks (DNNs). The paper highlights two SPS variants: one using the WORLD vocoder for faster, albeit slightly lower quality, synthesis, and another using WaveRNN for very high-quality but slower synthesis. For SPS, users can directly modify underlying model characteristics related to speed and pitch.

Beyond influencing the synthesis process, the tool also offers a post-processing editor. This allows for direct modification of the synthesized speech signal, including adjustments to speed, pitch, and volume.

Datasets

Not discussed in the paper. The paper mentions "large speech corpora" for unit selection and "training data" for SPS, but specific datasets are not detailed.

Key Insights and Ideas

  • Interactive Synthesis for Quality Tuning: The most significant insight is the development of a system that allows for "interactive speech synthesis." This means users can not only generate speech but also iteratively refine it by tweaking the synthesis parameters or directly editing the output.
  • Bridging Synthesis and Editing: The tool effectively combines a speech synthesizer with an audio editor, providing a unified environment for creating and polishing synthetic speech.
  • Subjective vs. Objective Optimization: For unit selection, the system acknowledges that the "best" sequence of units according to objective criteria might not always yield the most pleasant-sounding speech. The editor allows users to prioritize subjective quality by exploring alternative unit sequences.
  • Flexibility in Synthesis Methods: The system's ability to work with both unit selection and different SPS approaches (WORLD and WaveRNN) offers users a choice based on their needs for speed versus quality.

Key Results

The paper presents a functional web-based editor that allows users to:

  • Synthesize speech from general texts using multiple languages and voices.
  • Influence the synthesis process by modifying parameters for unit selection or SPS models.
  • Post-process synthesized speech to adjust speed, pitch, and volume.

While specific quantitative results (e.g., objective quality scores, speedup percentages) are not detailed, the paper emphasizes the tool's capability to improve subjective quality and fit specific requirements, particularly for pre-recorded announcements or similar applications.

Limitations and Future Work

  • Post-processing Quality Deterioration: The authors note that directly modifying the speech signal using post-processing tools can potentially deteriorate the quality of the synthesized speech.
  • On-line Synthesis Not Required: The tool is explicitly designed for scenarios where on-line synthesis is not a strict requirement, focusing on pre-prepared utterances.
  • Future Work: Not explicitly detailed in the provided text, but the nature of the tool suggests potential for further integration of more advanced editing features, broader support for synthesis methods, and perhaps more sophisticated objective metrics for guiding subjective improvements.

Summary updated

BibTeX
@inproceedings{gruber2019web,
  author = {Martin Grůber and Jakub Vít and Jindřich Matoušek},
  title = {{Web-Based Speech Synthesis Editor}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure