Suzune S1: An Ultra-Lightweight, High-Fidelity 80-Million Parameter Non-Autoregressive TTS Engine
Abstract
State-of-the-art neural Text-to-Speech (TTS) systems increasingly diverge into two suboptimal extremes: massive autoregressive (AR) audio foundation models that suffer from non-deterministic hallucinations, word skipping, high latency, and quadratically growing memory consumption; and conventional non-autoregressive (NAR) pipelines that rely on over-smoothed mel-spectrogram regression and computationally expensive multi-stage vocoding.
In this paper, we introduce Suzune S1, an ultra-lightweight, high-fidelity 80-million parameter non-autoregressive TTS engine engineered by Darsh Yadav at Sorika Labs for real-time edge and cloud deployment. Suzune S1 unifies phonemic language modeling via a customized 12-layer Phoneme-Level ALBERT (PL-BERT) contextual representation, an explicit continuous duration and prosody predictor with Adaptive Layer Normalization (AdaLN), and an end-to-end fast Inverse Short-Time Fourier Transform neural vocoder (iSTFTNet) powered by Neural Source Filter (NSF) harmonic excitation and periodic Snake activation functions.
By directly predicting multi-band STFT magnitude and phase components rather than time-domain raw sample regressions, Suzune S1 achieves broadcast-quality 24kHz audio synthesis at a Real-Time Factor (RTF) of 0.018 on modern GPUs and 0.087 on standard x86 CPU cores, maintaining a strictly constant ~320 MB VRAM footprint across unbounded inference sequences. We formalize the complete mathematical foundations, continuous duration projection equations, multi-resolution spectral loss landscapes, and style latent interpolation manifolds across bilingual (English and Hindi) acoustic corpora. Experimental evaluations show that Suzune S1 achieves a Mean Opinion Score (MOS) of 4.38 ± 0.06, rivaling commercial closed-source engines while reducing compute requirements by more than 85%.
Interactive 24kHz Acoustic Stream
Live iSTFTNet Inference • 80M Parameter Checkpoint1. Introduction and Architectural Motivation
Neural Text-to-Speech (TTS) synthesis has seen dramatic breakthroughs over recent years, transitioning from concatenative systems to deep generative models. Contemporary commercial demands increasingly require TTS systems capable of generating natural, expressive, broadcast-grade speech while operating with sub-100 millisecond latency and minimal compute consumption.
1.1 Limitations of Current Autoregressive Paradigms
Modern large discrete audio codec language models formulate speech generation as next-token prediction over acoustic codebooks. Despite high prosodic expressivity, discrete autoregressive models introduce catastrophic operational flaws in production environments:
- Phonetic Hallucinations and Word Skips: Due to unconstrained stochastic sampling, models occasionally skip syllables, repeat words, or hallucinate non-existent acoustic artifacts.
- Inference Latency Bottleneck: Autoregressive generation proceeds sequentially, resulting in a severe serial bottleneck where Time-To-First-Audio (TTFA) frequently exceeds 500ms.
- Dynamic VRAM Growth ($O(N)$): Key-value (KV) cache memory grows linearly with generated sequence length, causing Out-Of-Memory (OOM) errors in long-form generation tasks.
1.2 The Suzune S1 Solution
To overcome these challenges, Darsh Yadav at Sorika Labs engineered Suzune S1, an 80-million parameter end-to-end TTS engine that eliminates intermediate mel-spectrogram bottlenecks and avoids autoregressive decoding altogether.
2. Neural Architecture & Mathematical Formulation
Suzune S1 is composed of three interconnected sub-systems designed for end-to-end differentiability and parallel computation:
A. 12-Layer Phoneme-Level ALBERT (PL-BERT)
To overcome the semantic opacity of raw characters, we utilize a customized 12-layer PL-BERT with cross-layer parameter sharing. The phoneme encoder receives text converted to International Phonetic Alphabet (IPA) tokens and contextualizes phonemes bidirectionally.
B. Continuous Duration & Prosody Predictor (AdaLN)
Duration prediction is framed as a deterministic continuous regression problem. Adaptive Layer Normalization (AdaLN) dynamically modulates pitch ($F_0$), energy contours, and duration vectors conditioned on speaker style latent vectors.
C. Fast iSTFTNet Neural Vocoder with NSF
Instead of regressing 24,000 raw time-domain samples per second, the iSTFTNet vocoder directly predicts multi-band Short-Time Fourier Transform (STFT) magnitude and phase components. An integrated Neural Source Filter (NSF) provides harmonic sine excitation, while periodic Snake activation functions maintain vocal resonance.
3. Experimental Evaluation
We evaluated Suzune S1 on subjective audio quality (Mean Opinion Score, MOS), inference Real-Time Factor (RTF), and memory consumption across English and Hindi evaluation sets.
| Model | Params | MOS (1–5) | GPU RTF | x86 CPU RTF | VRAM | Word Skips |
|---|---|---|---|---|---|---|
| Suzune S1 (Ours) | 80M | 4.38 ± 0.06 | 0.018x | 0.087x | ~320 MB | 0% |
| FastSpeech 2 + HiFi-GAN | 45M | 3.84 ± 0.08 | 0.042x | 0.190x | ~280 MB | 0% |
| VITS Foundation | 110M | 4.21 ± 0.07 | 0.058x | 0.240x | ~480 MB | 0% |
| Autoregressive 1B+ Audio LM | 1.2B | 4.42 ± 0.05 | 0.420x | N/A | > 3.5 GB | Frequent |
Citation
@article{yadav2026suzune,
title={Suzune S1: An Ultra-Lightweight, High-Fidelity 80-Million Parameter Non-Autoregressive TTS Engine for Real-Time Edge and Cloud Deployment},
author={Yadav, Darsh},
journal={Sorika Labs Technical Report},
year={2026},
month={August},
url={https://sorikalabs.com/research/suzune}
}