ReadAloud TTS

ReadAloudHttpTTSService provides low-latency HTTP streaming speech synthesis with 8 kHz telephony output in Pipecat pipelines.

Overview

ReadAloud is a streaming text-to-speech API built for voice agents and phone calls. ReadAloudHttpTTSService streams audio from the ReadAloud HTTP endpoint as it is generated, so a pipeline can start speaking within a few hundred milliseconds. It outputs 24 kHz, 16 kHz or 8 kHz PCM and mu-law or A-law, which suits Twilio and Telnyx media streams. The integration is maintained by the author of ReadAloud, the speech API provider.

Source Repository

Source code, runnable example and changelog

PyPI Package

The pipecat-readaloud package on PyPI

Installation

uv add pipecat-readaloud

Last tested with Pipecat 1.12.0. The package requires pipecat-ai>=1.12,<2.

Prerequisites

• A ReadAloud API key from readaloudai.org/developers. New keys include a small free allowance.
• Set READALOUD_API_KEY in your application environment.

Configuration

Parameter	Default	Description
api_key	READALOUD_API_KEY env var	ReadAloud API key
base_url	https://api.readaloudai.org	API origin
engine	"piper"	"piper" (default, fastest) or "kokoro"
output_format	"auto"	auto, pcm_24000, pcm_16000, pcm_8000, mulaw_8000 or alaw_8000
sample_rate	Pipeline rate	Pipeline output rate. Use 8000 for Twilio and Telnyx
aiohttp_session	Created if omitted	Shared aiohttp.ClientSession
max_retries	3	Retries before the first audio byte on capacity, 5xx and connection errors

Additional keyword arguments pass to Pipecat’s TTSService. voice and speed are runtime settings: change them with ReadAloudHttpTTSService.Settings(voice="default", speed=1.1) inside a TTSUpdateSettingsFrame(delta=...).

Usage

import os
import aiohttp
from pipecat.pipeline.pipeline import Pipeline
from pipecat_readaloud import ReadAloudHttpTTSService

async def build_pipeline(transport, stt, llm):
    async with aiohttp.ClientSession() as session:
        tts = ReadAloudHttpTTSService(
            api_key=os.environ["READALOUD_API_KEY"],
            aiohttp_session=session,
            sample_rate=8000,  # 8000 for Twilio/Telnyx; 24000 is the native rate
            settings=ReadAloudHttpTTSService.Settings(voice="default", speed=1.0),
        )
        # Supply your application's transport, STT, LLM and context aggregators.
        return Pipeline([transport.input(), stt, llm, tts, transport.output()])

On interruption Pipecat cancels the in-flight synthesis and the connection is closed, which stops generation on the server. Capacity errors (HTTP 503 and 429) and 502/504 responses are retried with backoff until the first audio byte; authentication, quota and unknown-voice errors surface immediately as an ErrorFrame. The API returns no word timestamps, so TTSTextFrames are not time-aligned.

See the source repository for a single-file runnable example that needs only a ReadAloud API key, plus the changelog.