SABLE: a standard for TTS markup

Richard Sproat, Andrew J. Hunt, Mari Ostendorf, Paul Taylor, Alan W. Black, Kevin A. Lenzo, Mike EdgingtonICSLP 1998 · DOI

Abstract

In the late 1990s, the text-to-speech (TTS) industry faced significant fragmentation due to a lack of interoperability between proprietary synthesis engines. This paper introduces SABLE, an XML/SGML-based markup language designed to standardize the control of TTS systems. By providing a common syntax for both structural text description and speaker directives, SABLE enables developers to create speech-enabled content that is portable across different platforms and synthesizers.

Introduction

During the rapid development of speech technology in the late 20th century, the TTS landscape was characterized by a "Tower of Babel" effect. Major players such as Microsoft (SAPI), Apple (Speech Manager), and various academic research labs utilized incompatible, proprietary tag sets to control speech output. For developers, this fragmentation meant that an application built for one system required a complete rewrite of its markup to function on another. This lack of portability stifled innovation and discouraged investment in speech-enabled software. SABLE was proposed as a solution to this interoperability crisis, offering a single, extensible standard for the entire industry.

The SABLE Specification

SABLE is not a new synthesis algorithm but rather a robust specification for a markup language. It was developed by merging and refining two previous proposals: the Spoken Text Markup Language (STML) and the Java Speech Markup Language (JSML). The language is structured around two primary pillars:

1. Structural Text Description

These tags describe the nature of the text itself rather than its sound. For example, the DIV tag classifies regions as sentences or paragraphs. A critical component is the SAYAS tag, which informs the engine whether a string represents a date, a phone number, or a mathematical expression. This shifts the linguistic processing burden from the application developer to the TTS engine, ensuring that a date like "4/5/98" is interpreted correctly based on context (e.g., MDY vs. DMY).

2. Speaker Directives (Production)

These tags provide explicit instructions on the acoustic realization of the text. Key directives include:

  • EMPH: For adding emphasis to specific words.
  • PITCH, RATE, and VOLUME: For fine-grained prosodic control.
  • PRON: For specifying phonemic pronunciations using Unicode IPA.
  • AUDIO: For inserting pre-recorded audio files into the speech stream.

Key Design Principles

SABLE incorporates several architectural features to ensure longevity and flexibility:

  • Extensibility: To accommodate research and engine-specific features, SABLE allows for non-standard tags or attributes, provided they are prefixed with X- (e.g., <X-ME-PRON>). This prevents proprietary extensions from breaking the core standard.
  • Graceful Degradation: The specification dictates that if a TTS engine encounters an unrecognized tag, it should ignore the tag and synthesize the enclosed text. This ensures that the primary content remains accessible even if advanced formatting is lost.
  • Separation of Concerns: By distinguishing between structural elements and production elements, SABLE allows for "abstract specification." A developer can mark a boundary as a table row end, allowing the engine to determine the most natural auditory signal without requiring the developer to specify exact intonation patterns.

Results and Implementation

The authors successfully defined SABLE V0.2, creating a comprehensive framework that addresses the limitations of existing schemes like Microsoft’s SAPI or Apple’s Speech Manager. Unlike those systems, which focus almost exclusively on speaker directives, SABLE’s inclusion of structural text description allows for higher-quality, more natural-sounding speech.

Implementation of the standard is achieved through "translators"—software layers that convert SABLE tags into the native, idiosyncratic markup of legacy synthesizers. This provides a clear path for immediate adoption without requiring the total overhaul of existing synthesis engines.

Limitations and Future Directions

While SABLE represents a significant step forward, several challenges remain:

  • Language Fallback: Handling languages not supported by a specific engine is difficult; currently, systems may simply announce the language name rather than attempting synthesis.
  • Speaker Standardization: While SABLE supports gender and age attributes, standardizing specific voice names across different vendors remains a hurdle. Future iterations may move toward generic identifiers like MALE1 or FEMALE1.
  • Industry Adoption: The ultimate success of SABLE depends on its widespread adoption by commercial vendors and the academic community to further refine the Document Type Definition (DTD) and associated tools.

Conclusion

SABLE provides a vital bridge between incompatible TTS engines, offering a unified syntax that balances structural description with production control. By prioritizing portability and extensibility, it lays the groundwork for a more integrated and accessible speech technology ecosystem.

Summary updated

BibTeX
@inproceedings{sproat1998sable,
  author = {Richard Sproat and Andrew J. Hunt and Mari Ostendorf and Paul Taylor and Alan W. Black and Kevin A. Lenzo and Mike Edgington},
  title = {{SABLE: a standard for TTS markup}},
  booktitle = {ICSLP},
  year = {1998},
  doi = {10.21437/icslp.1998-14},
}
Figure