Piper

17 August 20265 min read

Piper is a fast, local neural text-to-speech (TTS) engine. It is designed for offline, low-latency speech synthesis on desktops, embedded devices, Raspberry Pi-class hardware, and home automation systems.

Current state

The original rhasspy/piper repository is archived and read-only. Development moved to the Open Home Foundation fork:

• old archived repo: https://github.com/rhasspy/piper

• current repo: https://github.com/OHF-Voice/piper1-gpl

So Piper is not “dead,” but the old repo is obsolete. The maintained line is OHF-Voice/piper1-gpl, published as piper-tts.

Also note the tradeoff: Piper is still useful for lightweight offline TTS, but it is not the most advanced neural voice generation system. For naturalness, voice cloning, emotional speech, or multilingual zero-shot voices, newer systems may sound better but usually need more compute and have more complicated licensing/serving concerns.

When to use Piper

Use Piper when you want:

• fast offline TTS;

• simple command-line synthesis;

• low-latency Home Assistant voice output;

• predictable local voices;

• a small TTS sidecar for an AI assistant;

• Raspberry Pi / edge-device speech output.

Use a newer/heavier TTS stack when you need:

• voice cloning;

• expressive/emotional speech;

• very natural long-form narration;

• multi-speaker conversational synthesis;

• commercial-grade studio quality.

Install

pip install piper-tts

For the HTTP server extras:

pip install 'piper-tts[http]'

Voices

List available voices:

python3 -m piper.download_voices

Download a voice:

python3 -m piper.download_voices en_US-lessac-medium

Voices are ONNX models plus JSON configs:

• en_US-lessac-medium.onnx

• en_US-lessac-medium.onnx.json

Voice quality levels are commonly named:

• x_low

• low

• medium

• high

Always check each voice’s MODEL_CARD/license. Piper code and Piper voice models are not the same licensing question.

CLI usage

# Write to a WAV file

python3 -m piper -m en_US-lessac-medium -f out.wav -- 'Hello world'

# Play immediately if ffplay is installed

python3 -m piper -m en_US-lessac-medium -- 'This will play on your speakers.'

# Use a custom voice directory

python3 -m piper --data-dir ./voices -m en_US-lessac-medium -f out.wav -- 'Hello from Piper'

Useful options:

• --cuda — enable GPU acceleration if onnxruntime-gpu is installed.

• --input-file — read text from one or more files.

• --sentence-silence — add silence between sentences.

• --volume — adjust output volume.

• --no-normalize — disable automatic volume normalization.

• --output-raw — stream raw audio bytes.

The CLI reloads the model each run, so for repeated use prefer the HTTP server.

HTTP server

python3 -m piper.http_server -m en_US-lessac-medium

This starts a server on port 5000 by default.

Open:

text

http://localhost:5000

Inspect voice/server info:

curl localhost:5000/info

curl localhost:5000/voices

Synthesize audio:

curl -X POST \

-H 'Content-Type: application/json' \

-d '{ "text": "This is a test." }' \

-o test.wav \

localhost:5000/synthesize

Request fields include:

• text — required text to synthesize.

• voice — optional voice name; defaults to server -m voice.

• speaker / speaker_id — for multi-speaker voices.

• length_scale — speaking speed.

• noise_scale — speech variability.

• noise_w_scale — phoneme width variability.

Python API

python

import wave

from piper import PiperVoice

voice = PiperVoice

.load("/path/to/en_US-lessac-medium.onnx")

with wave.open("output.wav", "wb") as wav_file:

voice.synthesize_wav("Hello from Piper!", wav_file)

CUDA:

python

voice = PiperVoice.load("/path/to/voice.onnx", use_cuda=True)

Streaming:

python

for chunk in voice.synthesize("Hello from streaming Piper"):

# chunk.sample_rate, chunk.sample_width, chunk.sample_channels

# chunk.audio_int16_bytes

write_raw_audio(chunk.audio_int16_bytes)

Home Assistant / Wyoming protocol

Piper is commonly used through Home Assistant’s Wyoming protocol.

yaml

services:

piper:

image: rhasspy/wyoming-piper

command: --voice en_US-lessac-medium

volumes:

piper_data:/data

ports:

- 10200:10200

Open Home Foundation also maintains voice-assistant-related projects such as wyoming-piper, wyoming-faster-whisper, and micro-wake-word.

AI assistant wiring

Piper is usually the TTS end of a local voice assistant:

flowchart TD
    Mic[Microphone]
    STT[Speech-to-text: whisper.cpp or faster-whisper]
    LLM[LLM/tool model through LiteLLM]
    TTS[Piper HTTP server]
    Speaker[Speaker]

    Mic --> STT --> LLM --> TTS --> Speaker

With LiteLLM, use Piper as a sidecar rather than trying to force it into the LLM gateway. LiteLLM routes model calls; Piper turns final text into audio.

Alternatives / newer TTS options

See also: Chatterbox for a stronger self-hosted TTS option when you want natural speech, voice cloning, multilingual generation, and voice-agent-friendly latency.

Piper remains good for lightweight offline TTS, but newer options may be better depending on goals:

Tool/model family	Best for	Tradeoff
Piper / `piper1-gpl`	Fast local deterministic voices, Home Assistant, edge devices	Less expressive/natural than newer large TTS systems.
Chatterbox	Advanced self-hosted voice cloning, multilingual TTS, and low-latency English agent speech	Heavier than Piper; GPU recommended.
CosyVoice	Production multilingual/cross-lingual streaming TTS	More complex deployment stack.
IndexTTS2	Emotion/style control and duration-control research direction	Newer stack; check exact feature availability by release.
Dia	Realistic two-speaker dialogue generation	English-focused and GPU-oriented.
Kokoro-style lightweight TTS	Small modern neural TTS with good quality for size	Ecosystem and licensing vary by implementation/model.
Coqui/XTTS-style systems	Voice cloning and multilingual expressive speech	Heavier; licensing and project maintenance need checking.
StyleTTS2 / research TTS	High-quality expressive generation experiments	More research-oriented; not as simple to deploy.
ElevenLabs / hosted TTS	Very natural hosted voices	External service, cost, privacy, and license constraints.
OpenAI/Azure/Google TTS	Hosted production APIs	Cloud dependency and provider pricing/policies.

Caveats

• The maintained Piper repo is GPL-3.0; check whether that matters for your distribution model.

• Voice models have their own licenses.

• The CLI is convenient but slow for repeated calls because it loads the model each time.

• Piper is excellent for simple offline TTS, but not necessarily state of the art for voice quality.