Tool calling

Data description

Scripted voice-agent calls: one human performs as agent while a second human performs as the caller. Agent fires mock tools during the call with realistic simulated latencies.

Each speaker’s microphone is captured locally as uncompressed 16-bit PCM WAV - never re-encoded from the call audio - and delivered as 24 kHz mono per-speaker stems (FLAC or WAV). Every take ships with the full scene script including that take’s rolled variable values, word-timed per-speaker transcripts keyed agent / caller, and a time-aligned events.jsonl logging every tool_call and tool_result against the delivered audio.

Each script is performed by multiple distinct speaker pairs, never the same speaker twice. Built for training and evaluating tool-calling voice agents: latency-bridging speech, turn-taking around function calls, read-backs and confirmations, and agent-persona TTS.