TTS services fixes.

[filipi87]

Summary

• Fixed word timestamp interleaving issue in ElevenLabsTTS when processing multiple sentences
• Added create_context_id() override to 6 TTS services to properly reuse context IDs across multiple run_tts invocations
• Fixed initialization logic to only emit TTSStartedFrame and start metrics once per turn, not once per sentence

The Problem

When an LLM generates multiple sentences in a single turn (before LLMFullResponseEndFrame), the base TTSService.create_context_id() creates a new UUID for each sentence. This caused:

1. Word timestamp interleaving - timestamps tracked in separate contexts instead of one continuous stream
2. Multiple TTSStartedFrame emissions - incorrect lifecycle signaling
3. Context tracking issues - base class tracking multiple contexts when only one is actively used

The root cause: frames are only paused when receiving LLMFullResponseEndFrame, but by that point run_tts() has been invoked multiple times (once per sentence). Some services need to reuse the same context_id across all invocations within a turn.

The Solution

Services that extend AudioContextTTSService or AudioContextWordTTSService and maintain self._context_id for turn tracking now override create_context_id() to:

1. Return existing self._context_id if one is already in progress
2. Create a new UUID only when starting a new turn
3. Reset self._context_id on interruptions or turn completion

Services Fixed

1. ElevenLabsTTSService - Context reuse + Word timestamp interleaving across sentences
2. InworldTTSService - Context reuse + Moved initialization inside guard
3. RimeTTSService - Context reuse
4. CartesiaTTSService - Context reuse
5. AsyncAITTSService - Context reuse + Moved initialization inside guard
6. PlayHTTTSService - Context reuse

Fixes #3723

[codecov]

Codecov Report

❌ Patch coverage is 0% with 49 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/pipecat/services/elevenlabs/tts.py	0.00%	19 Missing ⚠️
src/pipecat/services/asyncai/tts.py	0.00%	9 Missing ⚠️
src/pipecat/services/inworld/tts.py	0.00%	6 Missing ⚠️
src/pipecat/services/cartesia/tts.py	0.00%	5 Missing ⚠️
src/pipecat/services/playht/tts.py	0.00%	5 Missing ⚠️
src/pipecat/services/rime/tts.py	0.00%	5 Missing ⚠️

Files with missing lines	Coverage Δ
src/pipecat/services/cartesia/tts.py	0.00% <0.00%> (ø)
src/pipecat/services/playht/tts.py	0.00% <0.00%> (ø)
src/pipecat/services/rime/tts.py	0.00% <0.00%> (ø)
src/pipecat/services/inworld/tts.py	0.00% <0.00%> (ø)
src/pipecat/services/asyncai/tts.py	0.00% <0.00%> (ø)
src/pipecat/services/elevenlabs/tts.py	0.00% <0.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[markbackman]

Could we just move this logic to the base class?

[filipi87]

I don’t think we should for now. Most of the TTS services don’t save the context_id, and we already have a reference to it in each TTS service.

So maybe, in a follow up refactor, we can check which services are currently saving the context_id and refactor them. But for now, I would keep it like this.

[filipi87]

@markbackman, I’ve added this to our 1.0 wishlist, along with the other TTS improvements, so we can review how the classes are handling self._context_id. 👍

[markbackman]

LGTM.

This works as a temporary fix, but I agree that we'll want to make the API easier to work with in the future.