Piper TTS Review (2026): Fast, Local Neural Text-to-Speech

History: From Rhasspy to the Open Home Foundation

What Piper Actually Does

• Command-line synthesis. Run python3 -m piper with a downloaded voice model to synthesize a WAV file or stream audio directly to your speakers.
• Python API. Load a voice with PiperVoice.load() and call .synthesize_wav() or the streaming .synthesize() generator from your own Python application.
• HTTP web server mode. Piper can run as a persistent web server so voice models stay loaded in memory, avoiding the reload cost of the CLI on every call — recommended for repeated or production use.
• C/C++ API (libpiper). A native C++ library and CLI, ported from the legacy Piper repository, for embedding Piper in non-Python applications.
• Raw phoneme injection. Wrapping text in `[[ ... ]] lets you pass IPA phonemes (from espeak-ng --ipa=3`) directly, useful for correcting mispronounced names or technical terms.
• Optional GPU acceleration. Passing --cuda (CLI) or use_cuda=True (Python) enables CUDA acceleration via the onnxruntime-gpu package, though Piper is designed to run acceptably fast on CPU alone.
• Community-trained multilingual voices. Dozens of languages and regional variants are available as separately downloaded voice models, distributed via Hugging Face; voice quality varies by voice since they are trained by different community contributors.

Install and Run Piper: Step by Step

1. 1
   Install Piper.
   Why it matters: Run `pip install piper-tts` in a Python environment (Python 3.9+ recommended). This installs the `piper` package and its ONNX Runtime dependency; no GPU or CUDA setup is required for CPU-only use.
2. 2
   List and download a voice.
   Why it matters: Run `python3 -m piper.download_voices` with no arguments to list available voices, then `python3 -m piper.download_voices en_US-lessac-medium` to download a specific one to the current directory.
3. 3
   Synthesize speech from the command line.
   Why it matters: Run `python3 -m piper -m en_US-lessac-medium -f test.wav -- 'This is a test.'` to write a WAV file. If you have `ffplay` installed, drop the `-f` flag to hear the audio immediately instead of saving it.
4. 4
   (Optional) Use the Python API instead of the CLI.
   Why it matters: For repeated use inside an application, `from piper import PiperVoice; voice = PiperVoice.load("en_US-lessac-medium.onnx")` then `voice.synthesize_wav(text, wav_file)` avoids the CLI's per-call startup cost.
5. 5
   (Optional) Enable GPU acceleration.
   Why it matters: Install the `onnxruntime-gpu` package, then pass `--cuda` on the CLI or `use_cuda=True` to `PiperVoice.load()` in Python. This is optional — Piper is designed to run in real time on CPU alone.
6. 6
   (Optional) Run the web server for repeated use.
   Why it matters: For a production or repeated-use setup, run Piper's HTTP web server mode so the voice model stays loaded in memory instead of reloading on every CLI invocation.
7. 7
   (Optional) Fix a mispronounced word with raw phonemes.
   Why it matters: Get the IPA phonemes for a word with `espeak-ng -v en-us --ipa=3 -q <word>`, then wrap them in `[[ ... ]]` inside your input text to override Piper's automatic pronunciation for that word.

Real Usage Examples

• Raw phoneme injection for correcting pronunciation: wrap text like `The [[ bˈætmæn ]] not [[ bɹˈuːs wˈeɪn ]] using IPA phonemes obtained from espeak-ng -v en-us --ipa=3 -q <word>`.
• libpiper C/C++ API exposes piper_create() (or piper_create_with_options() for finer control) for embedding Piper directly in a native application without a Python runtime.

# Command line: write a WAV file
python3 -m piper -m en_US-lessac-medium -f test.wav -- "This is a test."

# Command line: play audio immediately (requires ffplay)
python3 -m piper -m en_US-lessac-medium -- "This will play on your speakers."

# Python API
import wave
from piper import PiperVoice

voice = PiperVoice.load("en_US-lessac-medium.onnx")
with wave.open("test.wav", "wb") as wav_file:
    voice.synthesize_wav("Welcome to the world of speech synthesis!", wav_file)

# Python API: adjust synthesis (speed, volume, expressiveness)
from piper import SynthesisConfig

syn_config = SynthesisConfig(
    volume=0.5,        # half as loud
    length_scale=2.0,  # twice as slow
    noise_scale=1.0,   # more audio variation
    noise_w_scale=1.0, # more speaking variation
)
voice.synthesize_wav("Custom synthesis settings.", wav_file, syn_config=syn_config)

# Python API: GPU acceleration (requires onnxruntime-gpu)
voice = PiperVoice.load("en_US-lessac-medium.onnx", use_cuda=True)

# Python API: streaming synthesis
for chunk in voice.synthesize("Streamed audio, chunk by chunk."):
    play_audio(chunk.audio_int16_bytes, chunk.sample_rate)

License and Cost

What Piper Is Not Good For

• Expressive, emotional, or few-second voice cloning. Piper synthesizes speech from pre-trained voice models, not from a short reference audio clip of a specific person. If you need to clone a voice from a few seconds of sample audio, or want more emotionally expressive delivery, XTTS v2 is built for that instead — see PromptQuorum's local TTS licensing guide for its (non-commercial) license terms.
• Multi-speaker cloning from a short sample. Similarly, Piper has no built-in mechanism to generate a new voice on the fly from an audio sample of a specific speaker; each voice is a separately trained and distributed model.
• GPL-3.0 copyleft obligations in a closed-source product. If your use case involves modifying and redistributing Piper's own source code inside a closed-source binary, the GPL-3.0-or-later license of the current repository is a real constraint that the original MIT-licensed repository did not have. Review the License and Cost section above and consult a lawyer before that kind of deployment.
• Guaranteed consistent voice quality across all languages. Because voices are trained and contributed by different community members, quality varies noticeably by language and by specific voice — check samples for your target language before committing to Piper for a production application.
• Long-term maintenance certainty. As of publication, the project's own README states the Open Home Foundation is looking for additional maintainers for Piper, which is worth factoring into a decision to build critical infrastructure on top of it.

Alternatives to Piper

Frequently Asked Questions

Verdict

Sources

• OHF-Voice/piper1-gpl on GitHub — the actively maintained repository: README, docs, license, and release history.
• piper-tts on PyPI — published package metadata, including the current GPL-3.0-or-later license.
• rhasspy/piper on GitHub — the original, now-archived repository (MIT license).
• piper1-gpl releases — version history, including v1.8.0 (September 4, 2026).
• Open Home Foundation — the nonprofit organization maintaining Piper and Home Assistant.

Related Reading

• Local TTS & Voice Cloning Licenses: Piper, XTTS v2, F5-TTS and Coqui — the licensing deep dive across all major local TTS engines.
• ElevenLabs vs Piper vs XTTS v2 — the cloud-vs-local comparison for readers deciding whether to self-host.
• Build a Fully Offline Voice Assistant in 2026 — connect Piper to a local LLM and a Whisper-based STT engine into a complete voice pipeline.
• Whisper.cpp Review — the local speech-to-text counterpart, reviewed in the same depth.
• faster-whisper Review — the CTranslate2-based speech-to-text alternative, reviewed in the same depth.