Cookbook: Voice Mode 1 (STT JSON)
Weight-free Deepgram/AssemblyAI/Whisper-segment converters — no raw audio in Gateway.
Cookbook — Voice Mode 1 (client STT JSON)
Convert vendor STT JSON to messages + voice_telemetry on the client. Run SequentialAnonymizer before evaluate. Do not send raw audio URLs to the Gateway (SSRF / PII / GPU rules).
from qurator import Qurator, voice
client = Qurator()
bundle = voice.from_deepgram(deepgram_json, anonymize=True)
# or: voice.from_assemblyai(...); voice.from_whisper_segments(segments_only)
result = client.evaluate(
messages=bundle["messages"],
criteria=["Disclose fees clearly; do not invent policy."],
voice_telemetry=bundle["voice_telemetry"],
assurance_tier="flash",
)Whisper segment JSON only — never load Whisper weights on the Mac or in Gateway for this path.