oto / research
Research for conversation.
We study the data, evaluations, and models needed for voice systems to participate in natural human conversation.
Latest research
otoSpeech Task: Task-oriented full-duplex dataset
Cascaded systems optimize task accuracy, full-duplex models optimize natural conversation, and both are heading for the same corner. That corner needs conversation data with the task still attached.
read the release →
otoSpeech is the training set for TurnBench
Sesame has released TurnBench, an open benchmark for turn-taking in spoken dialogue. Its 104-hour training set is otoSpeech, collected and hand-annotated by oto.
read the announcement →
otoTurn: End-of-turn detection for spontaneous conversation
Existing end-of-turn models work well for task-oriented voice agents. We tested what happens when the conversation stops behaving like a task.
read the research →