--- language: - tr license: apache-2.0 tags: - text-to-speech - tts - turkish - onnx - int8 - browser library_name: pytorch pipeline_tag: text-to-speech --- # zerre-tts (tr-v1) — an 8.5 MB Turkish TTS that runs in a browser tab A two-voice Turkish text-to-speech model, trained from scratch. The **int8 ONNX build is 8.5 MB** and runs in the browser via WebAssembly — no server, no API, no GPU. - 🎧 Demo: https://erkinemreta1.github.io/zerre-tts/ - 🧑‍💻 Code and training recipe: https://github.com/erkinemreta1/zerre-tts - Architecture adapted from [tiny-voca](https://github.com/n4ze3m/tiny-voca) (Apache-2.0). Türkçe: iki sesli, sıfırdan eğitilmiş Türkçe metin-konuşma modeli. Tarayıcıda çalışır, sunucu gerekmez; ölçümler ve kıyaslar aşağıdadır. ## Why this one | | | |---|---| | Size | **8.5 MB int8** (fp32: 21.4 MB) | | Parameters | 4.5M | | Voices | 24 kHz, two voices (SES-A, SES-B) | | Licence | Apache-2.0 — commercial use allowed | | Runs on | Browser (WASM), CPU, ONNX Runtime, PyTorch | | Turkish WER | **1.3%** (20 held-out clips) · **12.7%** (8 new sentences) | Quantisation, measured on 8 held-out clips with Whisper large-v3: | Variant | WER | File | |---|---:|---:| | fp32 | 14.0% | 21.38 MB | | **int8** | **15.3%** | **8.55 MB** | 13 MB smaller for no measurable cost in intelligibility. On the same clips, the post-processing chain (90 Hz high-pass, +2.5 dB presence at 3 kHz, spectral subtraction) halves the below-120 Hz energy that read as boom. ## Where it stands Among browser-capable Turkish TTS models, this one leads on size per unit of quality (eval sets differ; read the ratio, not the absolutes): | Model | Params | Deploy size | Turkish WER | MB/WER | Licence | Browser | Voices | |---|---:|---:|---:|---:|---|---|---| | **zerre-tts (int8)** | **4.5M** | **8.5 MB** | **1.3%** | **6.5** | **Apache-2.0** | ✅ | **2** | | okur-tts | 9.8M | ~23 MB | 1.81% | 12.8 | Apache-2.0 | ✅ | 1 | | Piper `tr_TR-dfki` | ~16M | 63 MB | 3.14% | 20.1 | CC BY-NC-SA | ✅ | 1 | | Meta MMS-TTS | 36M | ~145 MB | 5.74% | 25.3 | CC BY-NC | ✅ | 1 | | sanoTTS | 0.5-2.3M | ~2.8 MB | 30.3% | 0.09 | GPL-3.0 | ✅ | many | | Kokoro-82M | 82M | ~330 MB | no Turkish | — | Apache-2.0 | ✅ | many | | XTTS-v2 | 467M | ~2 GB | — | — | CPML (NC) | ❌ | cloning | Smallest deploy, best MB/WER, commercial-friendly licence, two voices. It is not the most natural: 1-2 GB server-class models still sound better. ## Files | File | Size | Use | |---|---:|---| | **`onnx/model-int8.onnx`** | **8.5 MB** | **Recommended.** Browser, embedded, distribution | | `onnx/model.onnx` | 21.4 MB | fp32, reference/comparison | | `onnx/model.json` | — | input/output spec, voice list, symbol table | | `inference.pt` | 18.2 MB | PyTorch checkpoint (retraining/fine-tuning) | | `samples/` | — | raw and post-processed examples | | `metrics/` | — | training journal, GPU utilisation report | ## Usage ### ONNX (int8) Inputs: `tokens` (int64 [1,N]), `stress` (int64 [1,N]), `voice` (int64 [1]), `speed`, `pitch_shift`, `pitch_range` (float32 [1]). Output: `audio` (float32). Text→phoneme is outside the graph: eSpeak NG `tr`, or the lexicon-plus-rules phonemizer used by the demo. ### PyTorch ```bash python -m zerre.infer --checkpoint inference.pt --lang tr \ --text "Merhaba dünya." --voice SES-A --output hello.wav ``` One sentence per call; split longer text at sentence boundaries. ## Training Openly licensed Turkish audiobook corpora (identities and the licence caveat: `docs/data_sources.md` in the repository): 16 h, downsampled 48→24 kHz, aligned with an eSpeak phoneme recogniser; 6,347 clips (11.93 h) accepted. 65,000 steps, ~6.5 h on a single L4, ~$7. ## Voice and data note The voices were trained from openly licensed recordings; **no cloning was done**, no speaker identity is claimed and there is no connection to the speakers. The voices are published under neutral labels (SES-A, SES-B). Details and caveats: `docs/data_sources.md`. Legal review for commercial use is the user's responsibility. ## Limits - Two fixed voices (SES-A, SES-B); no cloning, no streaming. - Read-speech prosody; questions and exclamations are weak. - Foreign names and abbreviations are often wrong. - WER was measured on small sets (20 clips / 8 sentences): a sanity check, not a benchmark. ## Citation ```bibtex @software{zerre_tts_2026, title = {zerre-tts: a 4.5M-parameter Turkish two-voice text-to-speech model}, year = {2026}, url = {https://github.com/erkinemreta1/zerre-tts} } ``` Architecture and training recipe: Muhammed Nazeem, *tiny-voca* (Apache-2.0).