Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model

Sahil Mahendrakar

cs.SDOct 5, 2026

Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model

Authors: Sahil Mahendrakar

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 them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher. It has 10x fewer parameters and needs 15x less compute. We first synthesize a corpus with the teacher and keep its durations, pitch, energy and phoneme features. We then train a small text side to predict these values, and a small decoder to turn the teacher's saved values into the teacher's audio, first with spectral losses and then adversarially. Finally, we connect the two halves and quantize the weights to int8. It needs no alignment learning and no joint training, and it runs on one laptop. Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 on UTMOS against the teacher's 4.52. The student initially kept a slight buzz, which we trace to the phase of voiced speech between 2 and 8 kHz. A phase-locking filter applied after synthesis removes most of it, with no training and no extra parameters. Code, model files and audio samples are at https://github.com/sahilmahendrakar/paradee

Figures & tables

Panel of Figure 1 — Figure 1: Kokoro’s two halves (top) and Paradee’s (bottom). Each student half learns from the frozen teacher separately (dashed arrows). At inference the two student halves are connected and turn phonemes into audio on their own, without the teacher.

Figure 2: Share of Kokoro’s parameters and of its compute (FLOPs per second of audio) in each part. The text side is ALBERT, the text encoder and the prosody predictor. The waveform generator is a quarter of the parameters but almost all of the arithmetic.

Subsystem	Parameters	Share
Decoder AdaIN encode/decode blocks (1024 channels)	33.6M	41%
Decoder iSTFTNet generator	19.7M	24%
LSTMs (text encoder and prosody predictor)	11.2M	14%
ALBERT (12 shared layers, 768 hidden) and its output projection	6.7M	8%
F0, energy and duration heads	6.6M	8%
Text encoder CNN and phoneme embedding	4.0M	5%

Table 1: Parameter census of Kokoro-82M v1.0.

	Kokoro-82M (teacher)	Paradee (student)
Parameters	81.8M	8.07M (10.1 $\times$ fewer)
Weights on disk	325 MB fp32 ONNX; 82 MB int8	8.45 MB int8
Compute per second of audio	55 GFLOP	3.4 GFLOP (15 $\times$ less)
Speed, PyTorch, 1 CPU thread	7.6 $\times$ real time	25.0 $\times$ real time
Decoder alone, onnxruntime, 1 CPU thread		20.5 $\times$ real time
UTMOS, fp32	4.52 $\pm$ 0.003	4.41 $\pm$ 0.01

Table 2: Paradee against its teacher. The student uses the 4.23M text side (feature-supervised, MLP projection), the 3.85M decoder after both training stages, and the phase-locking filter. UTMOS, WER and speaker similarity are over 200 held-out sentences, with 95% confidence intervals.

Model (voice)	Params	File size	Speed	UTMOS	WER
Kokoro-82M, teacher (af_heart)	81.8M	325 MB	7.6 $\times$	4.52 $\pm$ 0.003	5.7%
Paradee (af_heart)	8.07M	8.45 MB	25.0 $\times$	4.41 $\pm$ 0.01	5.7%
Paradee, released ONNX file	8.07M	9.0 MB	17.8 $\times$	4.41	6.0%
Kokoro-7M-Distill (af_msa)	7.48M	30.1 MB	35.5 $\times$	4.18 $\pm$ 0.03	7.4%
Piper, en_US-lessac-medium	15.7M	63.2 MB	15.4 $\times$	4.36 $\pm$ 0.02	8.8%
KittenTTS nano 0.8 (Bella)	14.0M	56.8 MB	10.7 $\times$	4.01 $\pm$ 0.04	5.5%

Table 3: Paradee against other small open models on the same 200 held-out sentences, with the same scoring. File size is the model as distributed (Paradee int8, the others fp32). Speed is real-time factor on one CPU thread of the same machine. Paradee and the Kokoro models are timed from phonemes. Piper and KittenTTS are timed from text, so their times include phonemization. Each model speaks its own voice, so UTMOS compares naturalness, not closeness to af_heart.

Text side, rendered by the teacher decoder	Params	DTW log-mel	UTMOS
Teacher text side	28M	0.12 (floor)	4.52
Student s, linear projection	4.0M	0.97	4.15
Student s, MLP projection	4.2M	0.91	4.36
Student m, MLP projection	7.1M	0.89	4.35
Student s, MLP, trained through the teacher decoder	4.2M	0.99	4.52

Table 4: Text-side students, each rendered by the frozen teacher decoder on the 200 held-out sentences. The teacher row’s DTW is the teacher against a second run of itself with a different random seed. The last row adds a log-mel loss on audio produced by passing the student’s outputs through the frozen teacher decoder. It has the worst DTW but the best UTMOS, because it learns prosody that sounds natural but differs from the teacher’s rendition.

Decoder A (3.85M) training	Decoder alone	Full student
Spectral losses only, 50k steps	2.98	3.00
+ adversarial, spectral weight 45, 3k steps	3.02
+ adversarial, spectral weight 10, 5k steps	4.29	4.32
+ adversarial, spectral weight 3, 5k more steps (stage two)	4.37	4.39
+ phase-locking filter (Paradee)		4.41

Table 5: UTMOS of the student decoder (driven by the teacher’s text side) and of the full student, over 200 held-out sentences. The teacher scores 4.52.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Reference point	DTW log-mel	Duration change
Teacher against itself, second run (random excitation)	0.09	0
All weights fp16	0.11	0
ONNX fp32 export against PyTorch	0.21	0
All weights int8 per channel, activations fp32	0.28	25 ms
Teacher at speed 0.97 (timing change only)	0.36	130 ms
Teacher on CPU against Apple MPS	0.39	0

Appendix

Table 6: Reference points for the distance metric.

Part	Params	8-bit	4-bit (groups)	4-bit	Timing shift, 8-bit
ALBERT	6.3M	0.263	0.495	0.555	10 ms
Text encoder	5.6M	0.153	0.266	0.319	5 ms
Prosody predictor	16.2M	0.312	0.379	0.402	25 ms
Decoder AdaIN	33.6M	0.090	0.126	0.179	0
Generator	19.7M	0.103	0.354	0.537	0
All weights	81.8M	0.270	0.624		25 ms

Appendix

Table 7: Quantizing the weights of one part of Kokoro at a time, with the rest at full precision (8-bit per channel, 4-bit in groups of 64, 4-bit per channel). Each value is the DTW log-mel distance from the full-precision teacher on 5 sentences. Lower is better, and 0.09, the teacher’s difference from its own second run, means no real change. The 0.4M projection after ALBERT is not quantized on its own. Snake activation parameters are kept in fp32, because quantizing them rounds small values to zero and produces NaNs.

Preset	Widths and kernels	Params	Generator	GFLOP/s audio
Teacher	1024 / 512 / 3,7,11	53.3M	19.7M	51.9
A (Paradee)	256 / 128 / 3,7,11	3.85M	1.19M	3.36
B	256 / 128 / 3,7	3.49M	0.83M	2.28
C	192 / 96 / 3,7	2.16M	0.48M	1.31
D	128 / 64 / 3,7	1.15M	0.23M	0.61
W2 (wider final stage)	128 / 256 / 3,7	4.39M	3.12M	8.2

Appendix

Table 8: Student decoders, each a copy of Kokoro’s decoder with scaled widths.

Explore similar work

DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech
Jul 5, 2026Junwon Moon, Seungbeom Kim, Yejin Lee +4Autoregressive Text-To-Speech Diffusion Language Models
1Sungkyunkwan University · University of Seoul
TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context
Sep 8, 2026Fritz Cremer, Jonathan CremerText-To-Speech Synthesis Acoustic Latent Space
FreyaTTS Technical Report
Jul 10, 2026Ahmet Erdem Pamuk, Ömer Yentür, Ahmet Tunga Bayrak +2F5-Tts Grapheme-To-Phoneme
Freya Team

Jul 5, 2026, eess.ASJ/K move · Enter open · S save

DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech

Junwon Moon, Seungbeom Kim, Yejin Lee, Hoseong Ahn+3

1Sungkyunkwan University · University of Seoul

Autoregressive Text-To-Speech Diffusion Language Models Autoregressive Generation Autoregressive Model

Open paper AskPDF arXiv

Autoregressive (AR) text-to-speech (TTS) models generate discrete speech tokens sequentially, which makes inference slow and can degrade robustness by propagating local errors and hallucinations. This limitation stems from their left-to-right AR commitment: each token must be determined before future speech-token context is available. However, such ordering is not an inherent requirement for TTS, as the full input text is available before synthesis. In this paper, we introduce DELTA-TTS, a lightweight LoRA-based adaptation framework that converts a pretrained AR TTS model into a discrete diffusion language model (dLLM) for confidence-ordered speech-token decoding. To better capture the local structure of speech, DELTA-TTS incorporates a convolution module that injects local acoustic context, together with a

1/t

-weighted training objective and a time-shifted inference schedule that defer low-confidence positions to later steps. Trained on only

585

hours of LibriTTS, DELTA-TTS achieves a

\textbf{1.75}\%

WER on Seed-TTS test-en, outperforming its AR backbone while generating tokens

\textbf{3.3}\times

faster. Further analysis shows that DELTA-TTS produces sharper text--speech alignment, increases overall decoding confidence, and mitigates hallucinations observed in AR generation.

Abstract

Explore similar work

DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech

TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context

FreyaTTS Technical Report