# How to Generate Speech Locally with Piper TTS

> Turn a text file into a WAV on your own machine — install Piper, download a voice, and choose between quality tiers and languages.

- Canonical: https://guides-ai.pages.dev/guides/text-to-speech-locally/
- Plate 13.05 · Topic: Everyday tasks (https://guides-ai.pages.dev/topics/tasks/)
- Published: 06 Sept 2026 · 3 min read
- Source site: guides-ai — https://guides-ai.pages.dev/

Piper is a small neural text-to-speech engine that runs on CPU — fast enough on a Raspberry Pi, instant on a laptop. It is the simplest local TTS to install: one pip package, one voice download, one command. Nothing is uploaded.

## 1. Install

```bash
pip install piper-tts
```

Installs the `piper` module and its ONNX runtime. No GPU required. On Windows use `python` instead of `python3` in every command below.

## 2. Download a voice

```bash
python3 -m piper.download_voices
python3 -m piper.download_voices en_US-lessac-medium
```

The first command lists every available voice; the second downloads one. Voices are named `<language>_<REGION>-<name>-<quality>`, where quality is `x_low`, `low`, `medium` or `high` — bigger is slower and better. Each voice is two files, an `.onnx` model and an `.onnx.json` config, saved **into the current directory** unless you pass `--data-dir <DIR>`.

Voices come from the `rhasspy/piper-voices` repository on Hugging Face and their licences differ — check the `MODEL_CARD` file next to a voice before you use it commercially.

## 3. Write a WAV

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

Writes `test.wav`. The `--` matters: it ends the flags so text starting with a dash is not parsed as an option. Drop `-f` entirely and Piper plays the audio through `ffplay` instead of saving it.

## 4. Read from a file

```bash
python3 -m piper -m en_US-lessac-medium --input-file chapter1.txt -f chapter1.wav
```

Synthesizes a whole text file into one WAV. Useful extras: `--sentence-silence 0.5` inserts half a second between sentences, `--volume 1.5` boosts the level, and `--cuda` uses an NVIDIA GPU if you also installed `onnxruntime-gpu`.

Piper reloads the model on every invocation, so a loop over hundreds of short strings is slow — run its HTTP server instead for that.

## 5. If the voice is not good enough

Piper trades naturalness for speed. **Kokoro** is the other easy local option: an 82M-parameter model with noticeably better prosody, still CPU-friendly.

```python
from kokoro import KPipeline
import soundfile as sf

pipeline = KPipeline(lang_code='a')          # 'a' = American English
for i, (_, _, audio) in enumerate(pipeline("Hello from a local model.", voice='af_heart')):
    sf.write(f"{i}.wav", audio, 24000)
```

Needs `pip install kokoro soundfile` plus the system package `espeak-ng`, and writes 24 kHz audio. All 54 voices ship inside the package.

---

Next: [go the other direction and transcribe audio](/guides/whisper-transcribe-audio-locally/) or [summarize a video before narrating it](/guides/summarize-youtube-video-with-ai/).