Alania-2: Turkish text-to-speech with instant and fine-tuned voice cloning

Sezgin Saygili, Emre Kaplaner, Oncel Ozgul, Fikri San KoktasarXiv 2026 · arXiv

TL;DR

Alania-2 is a 488M-parameter Turkish text-to-speech (TTS) system that achieves state-of-the-art intelligibility by prioritizing high-quality synthetic training data and robust text normalization over sheer model size. It delivers high-fidelity 48 kHz audio with a median first-audio latency of just 32 ms, featuring both instant voice cloning and fast-trained professional adapters.

What Problem Does This Paper Address?

Turkish is a "mid-resource" language for speech synthesis. While massive multilingual models exist, they are often optimized for English or Chinese, leaving Turkish with sub-par intelligibility.

The authors identify two primary hurdles:

  1. Linguistic Complexity: Turkish is an agglutinative language where numbers, dates, and foreign names take case suffixes that must agree with their pronunciation. Simple text-to-phoneme mapping often fails here.
  2. The "Size vs. Quality" Trade-off: There is a prevailing trend that "bigger is better" in TTS. The authors challenge this, seeking a model small enough for low-latency production (like telephone assistants) that doesn't sacrifice clarity.

Methods

Alania-2 utilizes the VoxCPM2 architecture, which consists of a language-model backbone that reads text and generates continuous audio latents via a diffusion transformer decoder. These latents are then turned into 48 kHz audio by a frozen Variational Autoencoder (VAE).

Overview of the Alania-2 architecture and pipeline.

Key technical innovations include:

  • Custom Text Normalizer: A rule-based engine that rewrites symbols and numbers into spoken words while handling complex Turkish vowel harmony for suffixes.
  • Optimized Inference: By implementing batched VAE decoding and CUDA-graph prefilling, the team reduced single-stream latency from 74 ms to 32 ms.
  • Two-Tiered Cloning:
    • Instant: Uses "reference-only" conditioning (rather than continuation) to prevent word skipping.
  • Professional: Employs rank-32 LoRA adapters trained in roughly two minutes to achieve high speaker similarity.

Datasets

The model was trained on a massive 18,195-hour corpus, though only 10% of this was "real" speech (filtered from YouTube/YODAS). The remaining 90% was synthetic speech generated by a larger teacher model and passed through rigorous quality filters.

The training data mix, highlighting the heavy reliance on high-quality synthetic speech.

The authors have released several resources to the community:

  • Freya-TR-Eval: A public benchmark of 495 Turkish sentences.
  • Captions & Text: A 2,398-hour caption set and a 1.44M-sentence domain text corpus.
  • Synthetic Speech Set: 3,411 hours of 48 kHz audio.

Key Insights and Ideas

  • Data Quality > Model Size: In a striking result, the authors trained "Antalia-2 Mini," a tiny 7.6M-parameter model, using the same data recipe. It matched the intelligibility of models 300 times its size. This suggests that for mid-resource languages, the bottleneck is often the training recipe, not the parameter count.
  • Normalization is Critical: Without the text normalizer, the model failed to correctly synthesize 100% of test sentences containing digits. With it, success jumped to 89.5%.
  • The "Continuation" Trap: Many zero-shot TTS models clone voices by "continuing" a prompt. The authors found this leads to instability (loops and skips) with longer references. Switching to "reference-only" mode (conditioning without transcribing the prompt) dropped word error rates from 21.9% to 1.7%.
  • Synthetic Teacher-Student Training: Using a large, slow model to generate "perfect" training data for a smaller, faster model is a highly effective strategy for languages where high-quality studio recordings are scarce.

Key Results

  • Intelligibility: On the Freya-TR-Eval benchmark, Alania-2 achieved a 1.44% Word Error Rate (WER), a massive improvement over their previous system's 5.99%.
  • Throughput: The model can handle 12 concurrent requests per second on a single H200 GPU while staying ahead of real-time.
  • Cloning Fidelity: Professional LoRA adapters significantly outperformed instant cloning in speaker similarity, with training taking less than 130 seconds on an A100.
Word Error Rate comparison across 15 different Turkish TTS systems.

Limitations and Future Work

  • Naturalness Metrics: The authors relied on UTMOS22, which is trained on English speech, potentially missing nuances of Turkish prosody.
  • Teacher Artifacts: Because 90% of the data is synthetic, Alania-2 occasionally inherits "robotic" high-frequency tones from its teacher model.
  • Safety: While professional clones require a live verification phrase, instant clones currently lack a robust "liveness" check, posing a potential risk for misuse.
  • Future Directions: The team plans to implement audio watermarking and conduct extensive subjective listening tests with native speakers.

Summary updated

Abstract

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.
BibTeX
@misc{saygili2026alania,
  author = {Sezgin Saygili and Emre Kaplaner and Oncel Ozgul and Fikri San Koktas},
  title = {{Alania-2: Turkish text-to-speech with instant and fine-tuned voice cloning}},
  year = {2026},
  eprint = {2611.13845},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13845},
}
Figure