MOS & TTS listening tests
Conduct subjective listening tests using MOS (Mean Opinion Score). Assess naturalness, pronunciation accuracy, prosody and voice consistency as required, and provide score summaries and examples of identified issues.
Model evaluation
We identify performance gaps through MOS listening tests, language quality assessment and multimodal evaluation. Task-specific independent test sets and consistent review criteria help guide model improvements.
Language & speech evaluation
Assess recognition, transcription and translation accuracy, and the relevance and factual accuracy of answers. Apply task-specific criteria and group issues by language and scenario.
Multimodal & preference evaluation
Review image understanding, image-text consistency and cross-modal alignment. Compare and rank responses to provide feedback data for RLHF.
Safety & adversarial evaluation
Use safety reviews and adversarial testing to analyze responses that violate safety boundaries, conflicting instructions and risk scenarios, and document issues for further review.
Let’s talk about
your data needs.
Share your data types, volume and expected timeline. We’ll discuss a specific plan with you.
- Address
- Room 3281, 2nd Floor, Building 2, No. 656-658 Wuyi Road, Changning District, Shanghai, China