Breaking

Sopro, a 135M lightweight TTS running on CPU, ships with voice cloning

June 29, 2026 at 10:05 EDT

Media Generation
Open Source

Samuel Vitorino has open-sourced Sopro, a lightweight English text-to-speech (TTS) model that runs at real time on a CPU. With 135M parameters, it supports zero-shot voice cloning and streaming generation, and is distributed under the Apache-2.0 license.

Feb 4, 2026 · Sopro TTS v1.5

A 135M text-to-speech model that clones voices on a CPU — faster than real time

An open-source, lightweight English TTS that does zero-shot voice cloning from just 3–12 seconds of audio, runs without a GPU, and was trained as a side project for roughly $100.

135M

Parameters (down from 169M in v1.5)

0.05

RTF on CPU — 32s of audio in ~1.77s

250ms

Time to first audio (streaming)

~$100

Training cost on a single L40S GPU

Speed: ~32 seconds of audio in ~1.77 seconds

RTF of 0.05 means generation runs ~20× faster than playback on a CPU (M3 basis).

32s

Audio produced

~1.77s

Compute time

A different architecture — built for size and speed, not SOTA quality

Heavy models

XTTS v2 · VoiceCraft

Transformer-based, large, need substantial GPU compute.

Sopro

WaveNet-style + light attention

Dilated convolutions, no Transformer stack — small and CPU-fast.

Praised for

Low cost, light footprint, easy to use
Near real-time generation with no GPU
Fits local voice assistants & edge devices

Trade-offs

Samples can sound "robotic" with artifacts
Below Chatterbox-TTS & Kokoro in quality
Hallucinates beyond the ~32s limit; English only

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.