active library Featured

narro

Narro: lightweight CPU text-to-speech

Started 2026 Python

Resources & Distribution

Source Code

GitHub Repository

Package Registries

Narro

Lightweight CPU text-to-speech. Wraps the Soprano-1.1-80M model for fast, expressive speech synthesis.

  • Up to 20x real-time on CPU
  • <1 GB memory, 80M parameters
  • 32kHz audio output
  • Infinite length via automatic text splitting

Installation

pip install narro

Or from source:

git clone https://github.com/ekwek1/soprano.git
cd soprano
pip install -e .

Usage

CLI

narro "Hello world" -o output.wav

Options:

--output, -o              Output audio file path (default: output.wav)
--model-path, -m          Path to local model directory (optional)
--no-compile              Disable torch.compile optimization
--no-quantize             Disable INT8 quantization
--decoder-batch-size, -bs Decoder batch size (default: 4)
--num-threads, -t         Number of CPU threads for inference

Note: The CLI reloads the model on each invocation. For repeated inference, use the Python API.

Python API

from narro import Narro

model = Narro(decoder_batch_size=4)

# Basic inference
out = model.infer("Hello world.")

# Save to file
out = model.infer("Hello world.", out_path="out.wav")

# Custom sampling parameters
out = model.infer("Hello world.", temperature=0.3, top_p=0.95, repetition_penalty=1.2)

# Batched inference
out = model.infer_batch(["Hello world."] * 10)

# Streaming inference (yields float32 tensors in [-1, 1])
for chunk in model.infer_stream("Hello world.", chunk_size=1):
    process_audio(chunk)

Usage Tips

  • Use double quotes instead of single quotes when quoting.
  • Best results with sentences between 2 and 30 seconds long.
  • Spell out numbers and special characters phonetically for best pronunciation (e.g., 1+1 -> one plus one).
  • Unsatisfactory results? Regenerate or adjust sampling settings.

License

Apache-2.0. See LICENSE for details.

Discussion