Multilingual Voice Questions Across 26 Locales

Simple Voice Questions (SVQ) is a collection of short audio questions recorded in 26 locales across 17 languages under multiple audio conditions. It contains approximately 700 speakers, with a cap of 250 recordings per speaker. The dataset was created by Google and serves as a core evaluation component for the Massive Sound Embedding Benchmark.

Use Cases

• Benchmarking automatic speech recognition models on audio questions recorded in 17 different languages.
• Evaluating model robustness across the four provided audio conditions: clean, background speech, media, and traffic noise.
• Analyzing speaker and gender class performance variations across the 26 locales.
• Training or fine-tuning speech embedding models on a multilingual corpus of short-form questions.

Strengths

• Covers 26 locales and 17 languages, providing broad linguistic diversity.
• Includes approximately 700 speakers, offering a substantial number of voice samples.
• Features recordings under four distinct audio conditions for robustness testing.

Limitations

• The total number of audio recordings (rows) is unknown.
• The dataset's specific license and file formats are not provided.
• The 'capped at 250 recordings per speaker' rule may limit data volume for some speakers.

Provenance

Source

Google

Collection Method

Recorded audio questions from speakers.

Time Range

null

Freshness

null

Geography

26 locales globally.