Piper Free code examples in C#, Python and JavaScript

THE LICENCE IS GPL-3.0, not MIT. Piper moved from rhasspy/piper to OHF-Voice/piper1-gpl in October 2025 and relicensed on the way. Check that against what you are shipping before it is load-bearing.

pip install piper-tts

# Voices are separate downloads, one ONNX file plus a JSON config.
python3 -m piper.download_voices en_US-lessac-medium

# Hindi, and everything else, works the same way.
python3 -m piper.download_voices hi_IN-pratham-medium

Useful for auditioning a voice and wrong for anything repeated: the CLI reloads the model on every invocation, so a loop over a thousand lines pays the load cost a thousand times.

python3 -m piper -m en_US-lessac-medium -f test.wav -- 'This is a test.'

# With ffplay installed, drop -f to hear it immediately.
python3 -m piper -m en_US-lessac-medium -- 'This is a test.'

PiperVoice.load() reads the model once. Hold that object for the life of the process — loading is the expensive part and synthesis is not.

import wave
from piper import PiperVoice

# Load once, at startup. This is the slow step.
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,
    )

SynthesisConfig is the whole tuning surface. length_scale is the one that matters: it is a MULTIPLIER on duration, so 0.8 is faster and 1.2 is slower — the opposite of what a field called speed would do.

from piper import SynthesisConfig

config = SynthesisConfig(
    length_scale=0.9,     # below 1.0 is FASTER
    noise_scale=0.667,    # audio variation
    noise_w_scale=0.8,    # speaking variation
    volume=1.0,
    normalize_audio=True,
)

with wave.open("faster.wav", "wb") as wav_file:
    voice.synthesize_wav(
        "The same sentence, read a little more quickly.",
        wav_file,
        syn_config=config,
    )

This is the shape worth deploying. One process holds the model, everything else makes an HTTP request, and no application needs onnxruntime as a dependency.

pip install 'piper-tts[http]'

python3 -m piper.http_server -m en_US-lessac-medium --port 5000

# Then, from anywhere:
curl -X POST --data 'This is a test.' http://localhost:5000 --output test.wav

No Node binding exists, and the GPL-3.0 licence is a reason to keep Piper behind a process boundary anyway. Run the HTTP server and fetch from it.

# On the box that will do the speaking:
pip install 'piper-tts[http]'
python3 -m piper.download_voices en_US-lessac-medium
python3 -m piper.http_server -m en_US-lessac-medium --port 5000

# Nothing to install on the Node side.

Plain text in, WAV bytes out. No JSON, no key, no account — the whole appeal of running it yourself.

import { writeFile } from "node:fs/promises";

const response = await fetch("http://localhost:5000", {
  method: "POST",
  headers: { "content-type": "text/plain" },
  body: "This is a test.",
});

if (!response.ok) {
  throw new Error(`Piper returned ${response.status}`);
}

await writeFile("test.wav", Buffer.from(await response.arrayBuffer()));

Piper has no CORS headers, so a browser on another origin cannot call it directly. Same-origin proxying is the fix, and it is the reason FreeYantra's TTS tool has a server route at all.

import express from "express";

const app = express();
app.use(express.text());

app.post("/api/speak", async (request, response) => {
  const upstream = await fetch("http://localhost:5000", {
    method: "POST",
    headers: { "content-type": "text/plain" },
    body: request.body,
  });

  if (!upstream.ok) {
    return response.status(502).json({ error: "synthesis failed" });
  }

  response.setHeader("Content-Type", "audio/wav");
  response.send(Buffer.from(await upstream.arrayBuffer()));
});

app.listen(3000);

Synthesis is CPU-bound and synchronous in the server. Concurrent requests compete for the same cores, so a queue in front of it is the difference between slow and unavailable.

let inFlight = 0;
const MAX_CONCURRENT = 2;     // roughly one per spare core

async function speak(text) {
  if (inFlight >= MAX_CONCURRENT) {
    throw Object.assign(new Error("busy"), { status: 429 });
  }

  inFlight++;
  try {
    const response = await fetch("http://localhost:5000", {
      method: "POST",
      headers: { "content-type": "text/plain" },
      body: text,
      signal: AbortSignal.timeout(5 * 60 * 1000),
    });

    if (!response.ok) throw new Error(`Piper returned ${response.status}`);
    return Buffer.from(await response.arrayBuffer());
  } finally {
    inFlight--;
  }
}