SonoroNova-ES: Large-Scale Synthetic English-to-Spanish Speech Translation

SonoroNova-ES is a large-scale synthetic English-to-Spanish speech-to-speech translation dataset containing 329,764 utterances. It was constructed via cascade pipelines combining text-to-text translation models with neural text-to-speech engines, using source audio derived from the HiFiTTS-2 English audiobook corpus. The dataset features 1,315 unique speakers and provides a total of 961 hours of audio.

Use Cases

• Train speech-to-speech translation models based on the described synthetic translation pipeline.
• Benchmark English-to-Spanish translation quality using the 961 hours of paired audio.
• Develop multi-speaker text-to-speech systems leveraging the 1,315 unique speaker voices.
• Study the quality and characteristics of synthetic speech data generated via cascade TTT and TTS models.

Strengths

• Large scale with 329,764 utterances.
• Includes audio from 1,315 unique speakers.
• Provides a substantial 961 hours of total audio duration.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect source bias inherent to the HiFiTTS-2 audiobook corpus.

Provenance

Source

HiFiTTS-2 English audiobook corpus.

Collection Method

Constructed via cascade pipelines combining open-weight text-to-text translation models with neural text-to-speech engines.

Freshness

Last updated 2026-05-07 00:15:15; freshness should be verified.