FreyaTTSA Compact Tokenizer‑Free Non‑Autoregressive Flow‑Matching Model for Turkish Speech Synthesis

We present FreyaTTS, an open, tokenizer-free Turkish speech model designed for reliable, efficient speech synthesis. FreyaTTS advances mid-resource TTS in three dimensions: (i) efficiency, packing competitive Turkish synthesis into a 183.2M-parameter model pretrained from scratch; (ii) Turkish-first, tokenizer-free design, driving generation from a 92-symbol character vocabulary with no phonemizer or discrete speech tokenizer, so that agglutinative morphology, vowel harmony, and in-context spoken-form numbers are learned end-to-end from audio; and (iii) non-autoregressive generation, via a conditional flow-matching Diffusion Transformer that predicts entire utterances in parallel inside the frozen continuous-latent space of AudioVAE2 (16 kHz encode, 48 kHz decode), with explicit cross-attention conditioning that aligns where E2/F5-style implicit alignment fails at this scale. On Freya-TR-Eval, a 495-sentence Turkish benchmark we release, FreyaTTS reaches 8.0% WER and 3.0% CER, and a 7-rater listening study places its naturalness above every same-class open system; on intelligibility it outperforms the larger open systems XTTS-v2 (~470M, 11.1%) and F5-TTS (~336M, 24.3%). Batched-ODE serving delivers 65.2 audio-seconds per second at batch 8 on one RTX 4090, 2.6 times the throughput of the 2B VoxCPM2 serving engine at a third of the memory, and the model runs at or faster than real time on a laptop CPU. We release the model weights, training and inference code, and the benchmark under the Apache-2.0 license.

All seven open sub-1B systems are evaluated on the 495-sentence Freya-TR-Eval set under one protocol: WER and CER from Whisper-large-v3 with 8 kHz band-matching, and naturalness from a listening study with 7 native Turkish raters (760 ratings, 95% confidence intervals).

Generation is one fixed 32-step solve with no KV-cache, so batching is nearly free: a batched-ODE server delivers 65.2 audio-seconds per wall-second at batch 8 on a single RTX 4090, 2.6 times the 2B VoxCPM2 serving engine at a third of the memory. The model also runs at or faster than real time on a laptop CPU.

FreyaTTS reads raw Turkish characters (92 symbols, no phonemizer), predicts how many 25 Hz latent frames the sentence needs, and denoises Gaussian noise of that length into speech latents in 32 Euler steps. A Diffusion Transformer aligns text and audio through cross-attention, and the frozen AudioVAE2 decodes the latents to a 48 kHz waveform. A fixed noise seed keeps the voice deterministic.

An interactive walkthrough renders the architecture as a navigable 3D scene, in English and Turkish: ten guided steps over the 92-symbol vocabulary, the ConvNeXt encoder, the duration MLP, one exploded DiT block, the 16-layer stack, and the 32-step Euler loop, with tensors at their true dimensions. Self-contained, no external requests, and a text fallback if WebGL is unavailable. Nothing loads until you press start.

The model is pretrained from scratch on multi-speaker Turkish, then fine-tuned in two short stages on one consented professional voice: a voice-lock stage that collapses cross-generation pitch spread from 74.9 to 5.0 Hz, and a short-utterance stage over forced-aligned one and two word segments. The released model is the final stage.

Freya-TR-Eval is 495 everyday conversational Turkish sentences with a fixed evaluation protocol: Whisper-large-v3 transcription with 8 kHz band-matching, Turkish-aware text normalization on both sides, and jiwer WER/CER. The benchmark and the evaluation code ship with the model under Apache-2.0: dataset at freyavoice/freya-tr-eval, harness at github.com/freyavoiceai/FreyaTTS.