We present Alania-2, a Turkish text-to-speech (TTS) system with two built-in voices, voice design from text descriptions, and two kinds of voice cloning. Alania-2 is a 488M-parameter model with the VoxCPM architecture, which generates continuous audio latents with a diffusion head on top of a language-model backbone. We train it from scratch on 18,195 hours of Turkish speech, 90% of which is synthetic speech produced by a larger teacher model and filtered automatically. A text normalizer rewrites numbers, abbreviations and foreign names into the words a Turkish speaker would say; without it, none of 96 test sentences with digits kept all its critical parts, such as an amount or a date, through synthesis and speech recognition, against 86 with it. On Freya-TR-Eval, a public benchmark of 495 Turkish sentences scored with automatic speech recognition, Alania-2 reaches a word error rate of 1.44%, compared with 5.99% for our previous system. On its built-in voices, it comes within half a point of character error rate of a 2.3B-parameter model at under a quarter of the size, and a 7.6M-parameter single-voice model trained with the same recipe matches the most intelligible systems we measured, which suggests that training data matters more for intelligibility than model size. We also report practical findings on voice cloning: conditioning on the reference audio alone, rather than continuing it, cuts skipped words from 21.9% to 1.7%; 20-30 seconds of reference audio works best; and a low-rank adapter trained in about two minutes raises speaker similarity well above instant cloning. Finally, we describe inference-engine changes that bring first audio down to 32 ms for a single stream. We release our evaluation harness and three Turkish speech and text datasets.
Figures & tables
Figure 1
Check
Measured with
Threshold
Intelligibility
CER, faster-whisper large-v3
≤ 0.05
Speaking rate
characters per second
7–25
Voice match
ECAPA similarity to the reference
≥ 0.45 ( ≥ 0.35 with a style instruction)
Naturalness
UTMOS22
≥ 2.5
Pitch
median F0
consistent with the voice’s gender
Table 1: Checks applied to every synthetic clip. A clip is kept only if it passes all of them.
Figure 2: Training data. The base mix is 90% synthetic speech; a short second stage adds voice design and style instructions.
Case
Written
Spoken form given to the model
Money with a suffix
750 TL'dir
yedi yüz elli liradır
Decimal amount
12,50 TL
on iki lira elli kuruş
Percentage with a suffix
%2,5'lik
yüzde iki buçukluk
Turkish initialism
SGK
segeka
Borrowed initialism
CEO
siio
Brand name
PatientDesk
Peyşınt Desk
Table 2: Examples of written input and the text Alania-2 receives.
As written
Respelled
Mean similarity, all 100 terms
0.911
0.961
Transcript equals the term or its respelling
72
85
Mean similarity, the 17 respelled terms
0.627
0.989
Table 3: Recognition of 100 brand and English terms by Duyu, in one carrier sentence with one voice and seed. Similarity is between the part of the transcript that fills the term’s slot and either the term or its respelling, whichever is higher (0–1) ↑ .
Figure 3: Choosing an architecture: intelligibility against naturalness on 120 Turkish prompts, 240 renders per system. The best systems sit at the top left.
Component
Layers × width
Heads
Params
Backbone LM
16 × 1024
8 (2 KV)
268M
Residual LM
6 × 1024
8 (2 KV)
72M
Feature encoder
8 × 768
6
69M
Diffusion decoder
8 × 768
6
72M
Total
488M
Table 4: Alania-2 architecture, excluding the frozen audio VAE. Attention heads stay 128-dimensional to match VoxCPM’s position encoding.
Figure 4: CER of the candidate models on our 292-clip suite (one seed; Whisper large-v3), lower is better. On its built-in voices, Alania-2 is close to the 2.3B models at under a quarter of their size, and the second training stage fixed most of the base model’s errors with style instructions. Full results are in Appendix A .
Figure 5: Freya-TR-Eval word error rate (495 sentences, three readings each, faster-whisper large-v3). Our models are in blue. Alania-2 through our API is seventh of the 15 systems; the bare model without serving checks is shown for reference. Antalia-2 Mini is our 7.6M-parameter single-voice model (Section 7 ).
Alania-1
Alania-2
Architecture
LLM over SNAC tokens
VoxCPM
Parameters
3.3B
488M
Sample rate
24 kHz
48 kHz
WER (%) ↓
5.99
1.44
CER (%) ↓
1.81
0.28
UTMOS22 ↑
3.19
3.67
Table 5: Alania-1 and Alania-2 on Freya-TR-Eval, both through our API. RTF is the real-time factor on one L4 GPU (Alania-1 measured on 100 sentences).
Figure 6: Freya-TR-Eval word error rate against model size for the 13 systems with published sizes (both axes logarithmic). Our models are in blue. Size does not predict intelligibility.
Figure 7: Streamed instant clones before and after the fix; lower is better. 200 streams each, with the same seeds and texts of 121–249 characters. “Hard” references contain reverberation, music, a laptop microphone, noise or MP3 compression.
Figure 13
Figure 9: Speaker similarity of professional clones (LoRA adapters) against instant clones of the same speakers, on held-out speech. Each row is one run; the last is the mean over four speakers through the serving engine.
Figure 10: Serving layout on each GPU machine.
Change
Metric
Before
After
Batched VAE decoding
VAE decode per step, 96 streams (ms) ↓
110
27
Audio seconds per second, 48 streams ↑
41
74
CUDA-graph prefill
First audio, 1 stream (ms) ↓
74
32
First audio, 96 streams (ms) ↓
143
94
Kernel warm-up
Stall on a new batch size (s) ↓
∼ 1.4
none
Table 7: Engine changes, measured on one A100 (40 GB) at the worker, excluding the network.
Figure 17
Path
Instant clone
Professional clone
Reference implementation
0.651 / 0.019
0.770 / 0.026
Our engine
0.659 / 0.019
0.772 / 0.028
Our engine, mixed batches
0.676 / 0.023
0.780 / 0.026
Table 8: Professional clones through the engine (4 speakers × 5 texts × 3 seeds). Speaker similarity ↑ / CER ↓ .
Antalia-2 Mini is a single-voice Turkish text-to-speech model with 7,620,931 parameters. It reads characters and produces 48 kHz audio at 41 times real time on an Apple M5 laptop CPU (2 threads). Its training audio came only from VoxCPM2 (the teacher), about 300 times larger, cloning one synthetic clip; automatic checks kept about 1,790 h. Fine-tuning the vocoder on the acoustic model's own mel spectrograms cut the word error rate (WER) of Whisper's transcripts from a mean of 3.89% to 1.75% at adjacent pre-release checkpoints. The teacher corrupts the start of a clip when the first word begins with ö, the student learned this, and a spoken prefix we cut from the output hides it. On the 495-sentence Freya-TR-Eval benchmark, where we tuned settings, version 1.0.1 has WER 1.09% and character error rate 0.22% on our public harness. A paired bootstrap cannot distinguish these from EMA Lightning, an 8.6M Turkish model, which scores higher on full-band UTMOS, an English-trained quality proxy. No human listening test exists. Weights, inference code and the benchmark harness are public under Apache-2.0; training code and data are not.
We present JaiTTS-v1.0, a state-of-the-art Thai voice cloning text-to-speech model built through continual training on a large Thai-centric speech corpus. The model architecture is adapted from VoxCPM, a tokenizer-free autoregressive TTS model. JaiTTS-v1.0 directly processes numerals and Thai-English code-switching, which is very common in realistic settings, without explicit text normalization. We test the models on short- and long-duration speech generation, which reflects many real-world use cases. JaiTTS-v1.0 achieves a state-of-the-art CER of 1.94%, surpassing the human ground truth of 1.98% for short-duration tasks while performing on par with human ground truth for long-duration tasks. In human judgment evaluations, our model wins 283 of 400 pairwise comparisons against commercial flagships, with only 58 losses. Our code and demo are available at https://github.com/JTS-AI-Team/JaiTTS .
We introduce Freya-TTS, a compact, tokenizer-free, Turkish-first text-to-speech model designed for highly reliable and efficient conversational synthesis. Freya-TTS is a 183.2M-parameter non-autoregressive conditional flow-matching Diffusion Transformer (DiT) that operates in the frozen continuous latent space of AudioVAE2 (16 kHz encode, 48 kHz decode), allowing the model to focus its capacity on text-to-latent mapping while inheriting high-quality 48 kHz reconstruction. We advance the framework along three key dimensions: (1) rule-free end-to-end modeling from a 92-symbol Turkish character vocabulary without a phonemizer, grapheme-to-phoneme frontend, or discrete speech tokenizer; (2) non-autoregressive parallel denoising, which predicts the entire latent sequence simultaneously over a predicted duration; and (3) a production-oriented two-stage post-training recipe consisting of single-speaker voice locking and short-utterance coverage, improving speaker consistency and robustness on short inputs. On the Freya-TR-Eval benchmark, Freya-TTS achieves a band-matched word error rate (WER) of 8.0% and character error rate (CER) of 3.0%, outperforming substantially larger open-source systems while using a fraction of their parameters. The model achieves a real-time factor of 0.11 on consumer GPUs and runs faster than real time on a laptop CPU, making it well suited for resource-constrained edge deployment. We release the model weights, training and inference code, and evaluation benchmark under the Apache-2.0 license.
Ahmet Erdem Pamuk, Ömer Yentür, Ahmet Tunga Bayrak +2