README.md

August 3, 2026 · View on GitHub

FreyaTTS-small: A Compact Tokenizer-Free Flow-Matching Transformer for Turkish-First Speech Synthesis

FreyaTTS-small is a 183M-parameter Turkish text-to-speech model, and the open-source member of the FreyaTTS family. It reads Turkish directly at the character level (92 symbols, no phonemizer, no G2P) and generates speech with a non-autoregressive conditional flow-matching DiT in the frozen AudioVAE2 latent space (25 Hz, 64-dim latents, 16 kHz encode / 48 kHz decode).

The result is a model that runs comfortably where 2B-class TTS models cannot: 1.5 GB of VRAM on a GPU, real time on a laptop CPU, and well under real time on the Apple Neural Engine.

Hear it first: freyatts.freyavoice.ai has 48 kHz audio samples, the benchmark numbers, and an interactive 3D walkthrough of the architecture.

The FreyaTTS family. FreyaTTS-small (this repository) is our compact, Apache-2.0, self-hostable model, released in full: weights, inference code, and training pipeline. FreyaTTS-large is our production model, serving Turkish voice agents at Freya (YC S25) with higher naturalness and expressivity; it is available commercially rather than as open weights. For access to FreyaTTS-large, contact us at dev@freyavoice.ai.

Repository and model identifiers (freyavoiceai/FreyaTTS, freyavoice/freya-tts) predate this naming and are unchanged; they refer to FreyaTTS-small.

Highlights

Small and fast - 183.2M parameters, RTF 0.10-0.11 and TTFT ~0.5 s on an RTX 4090, about 3.2x faster RTF and 3.7x less VRAM than the 2B VoxCPM2
Tokenizer-free Turkish - character-level input over a 92-symbol vocabulary, no phonemizer or G2P stage to maintain
Non-autoregressive - a single 32-step Euler ODE per clause, no classifier-free guidance needed
48 kHz output - the frozen AudioVAE2 decodes 25 Hz latents straight to 48 kHz audio
Runs on CPUs - real time on an Apple M3 laptop CPU (RTF 0.70 fp32), and RTF ~0.12 end to end via Core ML on Apple silicon
Apache-2.0 - weights and code free for commercial use

News

[2026.07] FreyaTTS-small 0.1.0 released: project page | weights | eval set

Quick Start

Installation

git clone https://github.com/freyavoiceai/FreyaTTS.git
cd FreyaTTS
pip install -r requirements.txt

Command line

python infer.py --text "Merhaba, size nasıl yardımcı olabilirim?" --out output.wav

Batch inference

Non-autoregressive generation batches naturally: one masked ODE solve serves a whole batch of requests. On an RTX 4090 this reaches about 65 seconds of audio per wall-second at batch size 8 in under 4 GB of VRAM.

python batch_infer.py --texts texts.txt --outdir wavs/ --batch-size 8

Python API

from freyatts import FreyaTTS

tts = FreyaTTS.from_pretrained("freyavoice/freya-tts", device="cuda")
wav = tts.synthesize("Merhaba, size nasıl yardımcı olabilirim?")   # np.float32, 48 kHz
tts.save_wav(wav, "output.wav")

from_pretrained also accepts a local directory containing config.json and model.safetensors. synthesize takes optional steps (default 32) and seed (default LEYLA_SEED) arguments.

The seed selects the speaker. FreyaTTS conditions only on text. There is no speaker embedding, speaker id, or reference-audio prefix, so the initial flow-matching noise x0 is the voice. Fine-tuning collapses the model onto the Leyla voice, but it does not condition on her, so a different seed gives a different person and some seeds are not Leyla at all. The default seed is pinned to the canonical Leyla voice and is deterministic: the same text returns the same waveform every call. Use one seed per utterance (the pipeline already shares it across clauses of a long input); passing seed=None opts into random-speaker sampling. Long inputs are normalized and split into clauses automatically, then joined with short pauses. Normalization spells out digit runs in Turkish (14:45 becomes on dört kırk beş): the duration predictor sizes the utterance from the character sequence, so numbers left as digits come out truncated.

Or run the bundled example:

python examples/synthesize.py --text "Merhaba, size nasıl yardımcı olabilirim?" --output output.wav

Model

FreyaTTS is a conditional flow-matching diffusion transformer (DiT) operating in the latent space of the frozen AudioVAE2:

Text encoding: character-level Turkish, 92 symbols shipped with the package (freyatts/char_vocab.json)
Generation: non-autoregressive flow matching, 32 Euler ODE steps, no CFG
Latent space: 64-dim latents at 25 Hz; AudioVAE2 encodes at 16 kHz and decodes at 48 kHz
Training: from scratch on Turkish speech; pretraining followed by SFT stage 1/2 (voice lock, short-utterance coverage)

AudioVAE2 is not retrained. It is downloaded from openbmb/VoxCPM2 (Apache-2.0) at load time via the voxcpm package.

Performance

Intelligibility (Freya-TR-Eval)

Evaluated on Freya-TR-Eval, a public Turkish TTS evaluation set.

Model	Params	WER ⬇	CER ⬇
FreyaTTS-small	183M	8.0%	3.0%
XTTS-v2	0.5B	11.1%	-
F5-TTS	0.3B	24.3%	-

FreyaTTS-small ranks 3rd of 7 among open sub-1B Turkish TTS models on this benchmark, at a fraction of the compute of the models above it. FreyaTTS-large is not included here; it is not an open model, and this table compares open sub-1B systems.

Speed

Setting	RTF	Notes
RTX 4090	0.10-0.11	TTFT ~0.5 s, 1.5 GB VRAM, 9.4 audio-s/s at concurrency 4
Apple M3 CPU (fp32)	0.70	real time on a laptop, no GPU
Apple silicon, Core ML	~0.12	end to end via the Neural Engine

Compared to the 2B VoxCPM2, FreyaTTS is about 3.2x faster in RTF and uses 3.7x less VRAM.

Reproduce with eval/benchmark.py (WER/CER) and eval/speed.py (latency/TTFT/RTF/concurrency).

Training

The training pipeline is included:

# encode audio to AudioVAE2 latents once, up front
python training/precompute_latents.py --manifest data/manifest.jsonl --output-dir data/latents

# pretraining
python training/pretrain.py --config training/configs/pretrain.yaml

# SFT stage 1/2 (voice lock, short-utterance coverage)
python training/sft.py --config training/configs/sft_stage1.yaml

License

FreyaTTS-small weights and code are released under the Apache-2.0 license. FreyaTTS-large is not covered by this license; contact dev@freyavoice.ai for commercial access.

Acknowledgments

VoxCPM2 (Apache-2.0) for the AudioVAE2 that FreyaTTS generates into; FreyaTTS reuses it frozen and unchanged
The flow matching and DiT literature this model builds on

Citation

FreyaTTS-small is described in our technical report, arXiv:2607.09530:

@misc{pamuk2026freyattscompacttokenizerfreeflowmatching,
      title={FreyaTTS: A Compact Tokenizer-Free Flow-Matching Transformer for Turkish-First Speech Synthesis}, 
      author={Ahmet Erdem Pamuk and Ömer Yentür and Ahmet Tunga Bayrak and Yavuz Alp Sencer Öztürk and Mustafa Yavuz},
      year={2026},
      eprint={2607.09530},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.09530}, 
}