Tazo Labs

Tazo Labs: Speech to Decision, Skip the Transcript

Tazo is an audio decision model for voice agents: it listens to raw call audio and picks the next action from a list of options you define, returning a calibrated probability for each. It runs as a standalone API next to an existing STT, LLM and TTS stack, so the LLM decides what to say while Tazo decides when, and what happens next.

Key Features

• Audio-native decisions: Tazo reads the call audio directly, with no transcript step, so it hears the pitch, pauses and tone that a transcript drops.

• Options you define: Each request carries a question and its options, such as respond or wait, keep talking or yield, backchannel, voicemail, escalate to a human, or start a tool call early. A new decision is a new option list, not a new model.

• Calibrated probabilities: Every option gets a probability whose confidence matches reality, so teams can set a threshold per action, for example escalating to a human only above 0.9.

• File and streaming modes: Upload a clip of up to 16 MB with POST /v1/decisions, or stream live PCM audio over a WebSocket and receive answers as the audio arrives.

• Shared call state: A request combines audio, a state field for context (string, object or list of strings) and a question with an ID that Tazo returns with the answer.

• Python SDK and agent skill: A Python SDK and an installable agent skill let coding agents set up an integration and verify decisions from audio.

• Stable v1 API: Fields may be added within v1, but existing fields and answer types are never renamed or removed; breaking changes ship as a new versioned endpoint.

• DECIDE benchmark: Tazo Labs publishes DECIDE, a benchmark of 4,867 decision questions over 1,081 clips from real production calls in English and Hindi, covering turn taking, graph routing, tool decisions, emotion and language switching.

Use Cases

• Turn taking: Decide whether a caller has finished or is still thinking before the agent responds.

• Call routing: Choose the next node in a graph-based voice agent, or hand the call over to a human.

• Early tool calls: Pick which tool to call straight from the audio, before the transcript and LLM turn complete.

• Call signals: Detect voicemail, caller emotion or a language switch during a live call.

Pricing

Usage is billed per second of audio processed, at $0.20 per hour of audio, through prepaid credits starting at $5. New accounts receive 15 minutes of free credits.

Getting Started

• Website: https://tazolabs.com

• Documentation: https://docs.tazolabs.com

• Benchmark: https://benchmark.tazolabs.com

Tazo gives cascaded voice agents a decision layer that works from the audio itself, letting teams keep their own models, prompts and tools while handing timing, routing and escalation calls to a model built for live conversations.