Sopro TTS

Sopro (from the Portuguese word for "breath/blow") is a lightweight voice-cloning text-to-speech model family. This repo ships sopro-v2-turbo, a 120M-parameter open model that streams, runs comfortably on a laptop CPU or in the browser, and reaches SOTA-level intelligibility against much larger systems. The full story, evaluations, and audio samples are in the blog post.

Main features:

120M parameters
English, European Portuguese, French, and German
Streaming with ~300 ms time-to-first-audio on a laptop CPU
Zero-shot voice cloning from 5-20 seconds of reference audio
0.24 RTF offline / 0.21 RTF streaming on an M3 CPU, 0.07 RTF on an H100
Runs in the browser via an ONNX runtime

Local demo

Run the model and open the demo with one command:

uvx --from sopro soprotts serve

If Sopro is already installed:

soprotts serve

Then navigate to http://localhost:7860. The model downloads on first use and stays in the Hugging Face cache. Sopro selects CUDA or CPU automatically and defaults to CPU on macOS (pass --device mps explicitly to use MPS). Use soprotts serve --help for model, device, port, and CPU int8 options.

Browser demo

There is also a fully in-browser demo with no server involved: https://samuel-vitorino.github.io/sopro/. On mobile the model is quantized, so results can be slightly below the local demo, and devices with low memory may crash. The dependency-isolated ONNX runtime and exporter are documented in web/README.md. Both demos use the frontend in demos/web.

Installation

From PyPI

pip install -U sopro

From the repo

git clone https://github.com/samuel-vitorino/sopro
cd sopro
pip install -e .

Examples

CLI

soprotts "Sopro is a lightweight 120 million parameter text-to-speech model that streams and runs on device." --ref ref.wav --out out.wav

Add --stream for the streaming path. You have the expected --temperature, --top-p, and --top-k parameters, alongside:

--lang (en, pt, fr, de; optional, helps pronunciation on ambiguous text)
--int8 (int8 AR weights on CPU)
--steps (acoustic solver steps; default 2)
--max-seconds (cap per generated segment; long text is split into segments, so total length is unbounded)

Python

Non-streaming

from sopro import SoproTTS

tts = SoproTTS.from_pretrained("samuel-vitorino/sopro-v2-turbo", device="cpu")

wav = tts.synthesize(
    "Hello! This is a non-streaming Sopro TTS example.",
    ref_audio_path="ref.wav",
)

tts.save_wav("out.wav", wav)

Streaming

import torch
from sopro import SoproTTS

tts = SoproTTS.from_pretrained("samuel-vitorino/sopro-v2-turbo", device="cpu")

chunks = []
for chunk in tts.stream(
    "Hello! This is a streaming Sopro TTS example.",
    ref_audio_path="ref.mp3",
):
    chunks.append(chunk.cpu())

wav = torch.cat(chunks, dim=-1)
tts.save_wav("out_stream.wav", wav)

You can also precalculate the reference to reduce time-to-first-audio:

import torch
from sopro import SoproTTS

tts = SoproTTS.from_pretrained("samuel-vitorino/sopro-v2-turbo", device="cpu")

ref = tts.prepare_reference(ref_audio_path="ref.mp3", stream=True)

chunks = []
for chunk in tts.stream(
    "Hello! This is a streaming Sopro TTS example.",
    ref=ref,
):
    chunks.append(chunk.cpu())

wav = torch.cat(chunks, dim=-1)
tts.save_wav("out_stream.wav", wav)

Disclaimers

We did not add watermarking: with an open-source inference pipeline it would be trivial to remove, so it would only provide a false sense of safety. Please use the model for good: do not impersonate people.
The text frontend is deliberately minimal, so some abbreviations, numbers, and symbols may not be pronounced correctly. Prefer words: 1 + 2 should be written one plus two. That said, Sopro generally reads common abbreviations like "CPU" or "TTS" fine, and you can put a language-specific normalizer in front of it.
Mixed-language text is a weak spot: words from one language inside a sentence of another (an English product name in a Portuguese sentence, for example) can be mispronounced.
The streaming path (chunked attention and the causal vocoder) is not bit-exact with the offline path. For best quality, prefer the offline path.
We are not planning to release the training code in the near future due to its complexity.

Training data

Acknowledgements

Downloads last month: -

Safetensors

Model size

0.1B params

Tensor type

F32