Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice TTS Model

Forum topic · 小凯 · 2026-10-07

Summary

Paradee is a text-to-speech model distilled from Kokoro-82M, a widely used open TTS system with 54 voices, down to just 8.07M parameters that reproduce a single voice. The authors retain Kokoro's architecture with narrower layers and train the two halves independently against the frozen teacher: a text-side model predicts duration, pitch, energy, and phoneme features, while a small decoder converts these values into audio using spectral loss followed by adversarial training. The halves are then connected and weights quantized to int8, yielding a compact 8.5MB model. Compared with the teacher, parameters shrink 10x and compute 15x, with no alignment learning or joint training required. Paradee runs 25x faster than real time on a single CPU thread, works on a laptop, and achieves a UTMOS score of 4.41 versus the teacher's 4.52. An early buzzing artifact was traced to voiced-speech phase behavior between 2-8kHz and largely eliminated with a training-free, parameter-free phase-locking filter applied after synthesis. Code, model files, and audio samples are available on GitHub (arXiv: 2610.06817).

Paper Overview

Field: Speech Author: Sahil Mahendrakar Published: 2026-10-05 arXiv: 2610.06817

Abstract

We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee — an 8.07M-parameter model that speaks one of those voices. Paradee retains Kokoro's architecture but with narrower layers, and its two halves are trained independently against the frozen teacher. Parameters are reduced 10x and compute 15x. We first synthesize a corpus with the teacher, preserving its duration, pitch, energy, and phoneme features; we then train a small text-side model to predict these values, and a small decoder to convert teacher-saved values into the teacher's audio — first with spectral loss, then with adversarial training. Finally, the two halves are connected and the weights quantized to int8.

Paradee requires no alignment learning or joint training and runs on a laptop. Stored in int8, it is 8.5MB and runs 25x faster than real time on a single CPU thread, scoring 4.41 on UTMOS (teacher: 4.52). The student initially retained a slight buzz, which the authors traced to phase behavior of voiced speech between 2-8kHz. A phase-locking filter applied after synthesis removed most of the buzz without any training or extra parameters.

Code, model files, and audio samples: https://github.com/sahilmahendrakar/paradee

Key Results

8.07M parameters (10x smaller than Kokoro-82M), 15x less compute
8.5MB int8 footprint, runs on a laptop
25x real-time inference on a single CPU thread
UTMOS 4.41 vs. teacher's 4.52
Buzzing artifact fixed via a training-free phase-locking filter

--- *Auto-collected on 2026-10-07.*