Written by Michał Kozłowski
on July 29, 2025

Text to speech in python

I’m working on integrating a speech synthesizer to my home assistant project. I decided to use piper library in python. I created a simple example of how to use piper python API to generate Polish audio from text. If you are interested in different languages the library supports over 40 of them (Check out those examples). This library does not need much computational power so it should work on edge devices like Raspberry PI or even better Nvidia’s Orin NANO. Latter supports cuda so will work even better.

Installation

Installation is very straight forward, you can download piper-tts pypi package using pip.

python -m venv .venv
source .venv/bin/activate
pip install piper-tts

then download the model

python3 -m piper.download_voices pl_PL-darkman-medium

you can quickly test everything

python3 -m piper -m pl_PL-darkman-medium -f test.wav -- 'To jest test!'

Streaming

There is no working example provided in the documentation, so I had to improvise. I want the assistant to be as responsive as possible, so streaming the response is important. I don’t want to wait until all of the text has been fully generated. The code provided below waits for the user to type in the message, and then reads it out to speakers.

import numpy as np
import sounddevice as sd
from piper.voice import PiperVoice, SynthesisConfig

# Setup
model = "./pl_PL-darkman-medium.onnx"
print(f"Loading voice model from: {model}")
voice = PiperVoice.load(model)
print("Voice loaded successfully.")

syn_config = SynthesisConfig(
            length_scale=1.2,  # Normal speed
            noise_scale=0.8,   # Moderate variation
            noise_w_scale=0.8, # Moderate speaking variation
            volume=0.5
        )

stream = sd.OutputStream(
    samplerate=voice.config.sample_rate,
    channels=1,
    dtype="int16",
)


stream.start()

text = input("> ").strip()
while text != "quit":
    for audio_chunk in voice.synthesize(text, syn_config=syn_config):
        raw_audio_bytes = audio_chunk.audio_int16_bytes
        # Convert the raw bytes to a NumPy array
        int_data = np.frombuffer(raw_audio_bytes, dtype=np.int16)
        stream.write(int_data)
    text = input("> ").strip()

stream.stop()
stream.close()
print("Program stopped")

Performance

I haven’t yet tested the performance on RPI or Orin but I will update this post as soon as I get to it. On my PC on the other hand without using cudathe response is instant. I’m satisfied with this result, other then quality of the voice there is no room for improvement.

→ Top