Tring

Open-source voice agent stack. One agent contract, any runtime, honest costs.

Tring is the sound of an arriving call. It lets you define a voice agent once and run it on any of three architectures — a cascaded pipeline (STT → LLM → TTS), a speech-to-speech model, or a hybrid of the two — without rewriting anything. It ships with a fully local, $0-per-minute pipeline as a first-class citizen, and it tells you exactly what every call costs.

Built from lessons learned running voice agents across hundreds of thousands of production telephony calls, in multiple languages, where callers code-switch mid-sentence and two seconds of silence means a hangup.

Why Tring

• One contract, three runtimes. AgentSpec is plain YAML. Switching a customer from a cascade to speech-to-speech is a config change: same tools, same analytics, same cost attribution.
• Local-first. faster-whisper for STT, Ollama for the LLM, Kokoro for TTS. Production-tuned turn-taking and interruption handling with zero cloud dependencies. Cloud providers are opt-in upgrades.
• Latency as correctness. In voice, dead air is a conversational error. Tring's primitives make it structurally impossible rather than patching it:
  • Tool-call choreography — the model cannot emit a tool call without declaring what to say while it runs (waiting_message, spoken_mode, post_tool_response are schema-required fields).
  • Speak-while-thinking — a character-level streaming parser flushes the speak field to TTS while the tool call is still being generated.
  • Interruption reconciliation — TTS generates faster than audio plays; on a barge-in, Tring annotates context so the model never references words the caller never actually heard.
  • Cache-safe language lock — the language directive is re-asserted every turn without breaking prompt-cache byte-stability.
• Honest cost accounting. Every metered unit carries an estimated flag. Per-call, per-connected-minute, per-conversation, and per-outcome roll-ups, so you optimize the denominator instead of the rate.
• Per-language provider routing. Real callers code-switch. The best STT engine for one language is often the wrong one for another; Tring routes STT/TTS/LLM per agent and per detected language.

Quickstart

pip install tring[local]

# agent.yaml
name: front-desk
persona: |
  You are a friendly front-desk assistant for a dental clinic.
language:
  primary: en
runtime:
  mode: cascade
  routing:
    default: { stt: faster_whisper, llm: ollama, tts: kokoro }

from tring import AgentSpec, CallSession
from tring.runtimes.cascade import CascadeRuntime

agent = AgentSpec.from_yaml("agent.yaml")
session = CallSession(agent)
runtime = CascadeRuntime(session)

See examples/ for the full local quickstart, the console dev loop, and cost reporting.

Status

Early alpha. The core contracts, primitives, cost meter, and local cascade runtime are usable; speech-to-speech and hybrid adapters, telephony ingress (SIP), and the hosted dashboard are on the roadmap.

Architecture

See docs/ARCHITECTURE.md.

License

MIT © Anzal Hussain Abidi