Transformation

Sometimes the LLM’s text needs a touch-up before it’s spoken: expand an acronym so TTS says it correctly, strip markdown, or filter sensitive content. The tts pipeline hook gives you the reply’s text stream so you can transform it before synthesis.

The `tts` hook

Register an async hook with @pipeline.on("tts"). It receives the streaming reply text; yield the transformed text, then hand it to run_tts() to produce the audio.

import re
from zrt import run_tts

pronunciation_map = {"nginx": "engine x", "API": "A P I", "SQL": "sequel"}

@pipeline.on("tts")
async def tts_node(text_stream):
    async def text_phase():
        async for response in text_stream:
            processed = response
            for word, say_as in pronunciation_map.items():
                processed = re.sub(rf"\b{word}\b", say_as, processed, flags=re.IGNORECASE)
            yield processed

    async for audio_chunk in run_tts(text_phase()):
        yield audio_chunk

What you can do in the hook

Fix pronunciation of acronyms, brand names, and technical terms (shown above).
Filter or redact content before it’s spoken.
Normalize formatting: strip markdown, expand numbers/dates, clean up symbols.

Because the hook sits between the LLM and TTS, the change affects only what’s spoken: the text recorded in chat context is unchanged.

This is one of several pipeline hooks. The same mechanism lets you intercept other stages of the pipeline.

References

Python

Examples

Enhanced Pronunciation

Rewrite text before synthesis.

Chat Overview

Get started

Modalities

Configure Pipeline

Tools and Capabilities

Turn Detection and Interruptions

Telephony Agents

Context Management

Observability and Monitoring

The `tts` hook

What you can do in the hook

References

Examples

Enhanced Pronunciation

​The tts hook

​What you can do in the hook

​References

​Examples

Enhanced Pronunciation

The `tts` hook

What you can do in the hook

References

Examples