piper

Fast local neural text-to-speech engine

TLDR

Synthesize speech

$ echo "Hello world" | piper --model [en_US-lessac-medium] --output_file [output.wav]

List models

$ piper --list-models

Use specific model

$ piper -m [model.onnx] -c [model.json] < [input.txt] > [output.wav]

Set speaker

$ piper --model [model.onnx] --speaker [0] < [input.txt]

Adjust speaking rate

$ piper --model [model.onnx] --length_scale [1.5] < [input.txt]

JSON input mode

$ echo '{"text": "Hello"}' | piper --model [model.onnx] --json-input

SYNOPSIS

piper [--model file] [--outputfile file] [options_]

piper is a fast, local neural text-to-speech system that generates natural-sounding speech from text using ONNX-based voice models. It runs entirely offline after model download, requiring no internet connection or cloud API.
Each model is trained for a specific language and voice. Multi-speaker models support selecting different voice variants via speaker ID. The lengthscale and noisescale parameters control speaking rate and variation, allowing fine-tuning of output characteristics.
Input is read from stdin and output as WAV audio. JSON input mode enables structured text processing with per-utterance settings. Models are available for many languages through the Piper project's model repository.

PARAMETERS

-m, --model FILE

ONNX model file.

-c, --config FILE

JSON config file.

--output_file FILE

Output WAV file.

--output_dir DIR

Output directory.

--speaker ID

Speaker ID for multi-speaker.

--length_scale FLOAT

Speaking rate (higher=slower).

--noise_scale FLOAT

Variation in speech.

--json-input

JSON input format.

--list-models

Show available models.

CAVEATS

Requires model download. Quality varies by model. GPU acceleration optional.

HISTORY

Piper was created by Michael Hansen (rhasspy) for offline voice assistants. It provides fast, high-quality TTS suitable for embedded and edge devices.