The speechdat-car multilingual speech databases for in-car applications: some first validation results

van den Heuvel, Henk; Boudy, Jerôme; Comeyne, Robrecht; Euler, Stephan; Moreno, Asuncion; Richard, Gael · 1999 · OpenAlex-citations

DOI: 10.21437/eurospeech.1999-574x

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

The SpeechDat-Car project addresses the challenge of developing robust multilingual speech recognition systems for in-car environments, where noise from engines, traffic, and audio equipment significantly degrades performance. Motivated by the need for hands-free interfaces to control car accessories and access teleservices without distracting drivers, the project aims to create a set of uniform, coherent speech databases for nine European languages: Danish, English, Finnish, Flemish/Dutch, French, German, Greek, Italian, and Spanish. Each database is designed to support both training and testing of speech recognizers, containing 600 recording sessions from at least 300 speakers balanced by age, gender, and regional accent. The data includes isolated words, navigation terms, digits, dates, phonetically rich sentences, and spontaneous speech, recorded under seven distinct environmental conditions ranging from stationary engines to high-speed highway driving with radio noise. The data collection utilizes a dual-platform recording system comprising a mobile platform (PltM) installed in the car and a fixed platform (PltF) connected via GSM. PltM records multi-channel speech at 16 kHz using one close-talk and three far-talk microphones, while PltF captures the transmitted GSM signal at 8 kHz. A strict validation protocol ensures data quality and exchangeability, consisting of platform validation (expert tests for audio quality, load, interrupt recovery, and stability) and functional tests involving user questionnaires. Database pre-validation involves submitting a mini-database of six sessions to detect design errors before full-scale recording. Initial validation results reported in April 1999 indicate that three databases successfully passed platform validation, while one was submitted for pre-validation. The low number of completed validations was attributed to unexpected delays in platform installation. Functional tests revealed that recording sessions lasted approximately 45 minutes. While participants generally appreciated the procedure and expressed willingness to participate again, many perceived the session length as long. Technically, the mobile platform reliably recorded all items even after interruptions, whereas the fixed platform occasionally missed one or two items due to temporary GSM disconnections. The significance of this work lies in the creation of high-quality, homogeneous speech resources that enable comparative research on automatic speech recognition across multiple languages in adverse automotive conditions. By standardizing the database design and validation procedures, SpeechDat-Car facilitates the development of robust voice-driven services and ensures that the resulting datasets are suitable for rigorous scientific evaluation. The project, funded under the 4th EC Framework, planned to deliver all nine databases for final validation by October 2000, providing the speech technology community with unique tools for advancing in-car speech processing.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• voice interaction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: dataset resource