[tomjad]

Summary

Adds Techmo TTS to the package, next to the existing ASR plugin:

from livekit.plugins.techmo import STT, TTS

session = AgentSession(stt=STT(language_group="pl"), llm=..., tts=TTS(voice_name="ewa"))

The code is a port of the internal Techmo plugin livekit-plugins-techmo-tts 0.3.2, used so far alongside this package. It goes in as the livekit/plugins/techmo/tts.py module of the single livekit-plugins-techmo distribution (the same layout as the official plugins, e.g. livekit.plugins.google.STT/.TTS). ASR users need no change, and the internal import path from livekit.plugins.techmo.tts import TTS keeps working.

What TTS provides

• native streaming synthesis (TTS.stream(): sentence by sentence, no StreamAdapter) and TTS.synthesize()
• voice / language / prosody options, update_options(), list_voices(); without voice_name / language_code the service's configured defaults are used (the internal plugin forced "michal" / "pl-PL"); prewarm() warns when the voice name is not offered by the service, since the service silently falls back to its default voice
• SSML <mark/> markers as MarkerEvent (on_marker callback)
• TLS / mutual TLS from bytes or file paths
• gRPC channel pool with lifetime rotation, prewarm(), reconnect on transient UNAVAILABLE; gRPC errors mapped to APIConnectionError / APITimeoutError, so livekit-agents' retry logic applies
• session_id sent as session_id gRPC metadata; not sent when unset (the internal plugin sent a random UUID). This matches tts-client-python without --session-id and the ASR plugin

Changes

File	Change
livekit/plugins/techmo/tts.py	Ported module. Logic unchanged: the AST is identical to the source except for the logger (child livekit.plugins.techmo.tts of the plugin logger), the user-agent (livekit-plugins-techmo/<version>) and the missing-stubs hint (git URL, since tts-api is not on PyPI). Long strings re-wrapped for the 120-column limit.
livekit/plugins/techmo/__init__.py	Exports TTS, TTSOptions, TechmoSynthesizeStream, MarkerEvent, VoiceInfo
pyproject.toml	tts-api @ git+…@v3.2.1; livekit-agents>=1.2 (was >=1.0); description, keywords
tests/test_tts.py, tests/test_tts_integration.py	100 unit tests + integration tests (skipped without TECHMO_TTS_ADDRESS); patch() targets moved to the new module path
.github/workflows/ci.yml	New min-deps job: unit tests with livekit-agents==1.2.0
README.md	Split into ASR and TTS sections (TTS configuration, prosody, TLS, markers, warm-up, session ID); requirements updated
KNOWN_ISSUES.md	Workarounds for Techmo TTS services older than 3.2.4 (internal tracker links removed)
examples/voice_agent.py	Voice agent with Techmo ASR + TTS (AgentSession API)
CHANGELOG.md, Makefile	Unreleased entry; make test skips integration tests, new make test-integration

Why livekit-agents>=1.2

TTS splits text into sentences with tokenize.blingfire, which livekit-agents 1.0 and 1.1 don't have: on 1.1.7, 70 tests fail. On 1.2.0 all pass. The new CI job keeps that floor tested; the matrix jobs test the newest version.

Dependencies

tts-api 3.2.1 ships stubs generated with grpcio 1.70 / protobuf 5.29, the same floors techmo-asr-api already requires, so grpcio>=1.70.0 / protobuf>=5.29.0 stay as they are. No file is shared between tts-api and techmo-asr-api (checked against both RECORDs) and, unlike techmo-asr-api, tts-api does not install anything under google/.

Release

This PR also bumps the version to 0.4.0 (pyproject.toml, version.py, CHANGELOG [0.4.0] - 2026-10-23), as was done for 0.2.0 in #1. After merge: tag 0.4.0 on the merge commit and create the GitHub release from the CHANGELOG section. If the merge slips by more than a day or two, update the date in CHANGELOG first.

Verification

• ruff, ruff format, mypy (5 files): clean
• unit tests: 129 passed (100 TTS + 19 ASR) with livekit-agents 1.2.0, 1.5.1 and 1.8.8, Python 3.10 and 3.12; clean editable install, as in CI
• uv build: wheel contains tts.py; Requires-Dist includes tts-api
• examples/voice_agent.py imports and constructs STT/TTS on livekit-agents 1.2.0 and 1.8.8
• Live Techmo TTS service (3.2.x, 210 voices): integration tests 8/8. A request built by the plugin without a voice gives the same audio as a request without synthesis_config (the service default); an unknown name such as MASZA triggers the prewarm() warning with Masza, masza as suggestions. Synthesis also checked from testenv.

Open points for review

• Session-ID metadata key: TTS sends session_id (underscore), ASR sends session-id. Confirmed against the reference client: tts-client-python uses ("session_id", …) and sends it only when --session-id is given (tts_client_python/general.py), so the keys stay as they are.
• examples/basic_stt.py (existing) still uses the pre-1.x VoicePipelineAgent API, which livekit-agents 1.x no longer has. Not changed here.
• README links to LICENSE, but the repository has no LICENSE file. Not changed here.

🤖 Generated with Claude Code