- Provider-reported duplex scores differ from multi-turn task results; they do not establish universal superiority.
A comparison of interaction skills
A November 13 analysis on OpenMOSS compares MOSS-Live with ByteDance’s Doubao SeedDuplex and OpenAI voice models using results published by the MOSS product provider. The article focuses on interruptions, irrelevant speech and the distinction between natural conversation and completing tasks.
In the reported FDB v1.5 evaluation, MOSS-Live’s irrelevant-speech false-response rate was 2%, versus 75% for Doubao SeedDuplex. Its FDB v3 tool-call success rate was 51%, tied with Doubao. These are attributed benchmark results, rather than independently reproduced measurements.
Task performance and disclosure limits
The comparison also preserves weaker results. On τ-Voice multi-turn tasks, MOSS-Live scored 57.93%, below GPT-Live-1’s 67.68%, although above Doubao’s 17.38%. Its Big-Bench Audio accuracy was 88.40%, below Doubao’s 90.50%.
The article describes a foreground conversation layer working alongside background reasoning and tool execution. It does not establish that this design improves every use case. The page does not disclose sample sizes, uncertainty ranges, API snapshots or complete evaluation settings, and warns against combining different benchmarks into an overall score. The report therefore supports a bounded comparison with a Chinese AI product, not a universal ranking or a claim of independent validation.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01


