Azerbaijani TTS — 9.36M, 24 kHz, offline

A text-to-speech model that speaks Azerbaijani entirely offline: no server, no API key, no network at inference. 9.36M parameters, 24 kHz mono, single speaker, 2-4x faster than real time on a laptop CPU.

It started as a fine-tune of owensong/Inflect-Micro-v2 by Owen Song. This is the second Azerbaijani checkpoint, v2.1: it continues from the first on 23,088 clips distilled from a teacher TTS, reads stress from a stress frontend on top of eSpeak NG, and has a decoder polished with stronger discriminators.

This checkpoint is the Azerbaijani voice of Avaz, an open-source text-to-speech project that runs on the device itself — in the browser, on a CPU, with no server — and aims at natural intonation on phones. Turkish and Russian are next.

Hear it The playground — runs in your browser, nothing installed
Code github.com/HuseynliIlqar/Inflect_Micro_v2_Azerbaijan
In this repo PyTorch at the root, ONNX in onnx/, WAVs in samples/

Samples

Text Audio
Salam, bu model tamamilə yerli maşında işləyir.
Payız gəlmişdi və şəhərin küçələri saralmış yarpaqlarla örtülmüşdü.
Sən bu kitabı oxumusan? Mənə çox maraqlı gəldi.
II Dünya müharibəsi 01/09/1939 tarixində başladı və 25% artım oldu.

The last one shows why the text layer matters: II becomes İkinci, 01/09/1939 becomes bir sentyabr min doqquz yüz otuz doqquzuncu il, 25% becomes iyirmi beş faiz.

Usage

The weights are useless without the Azerbaijani text layer, which lives in the GitHub repository together with the CLI, the tests and the training pipeline:

git lfs install
git clone https://github.com/HuseynliIlqar/Inflect_Micro_v2_Azerbaijan
cd Inflect_Micro_v2_Azerbaijan
pip install -r requirements.txt
python say.py "Salam, necəsiniz?"

That clone already contains these weights. To use the copy in this repository instead:

from huggingface_hub import snapshot_download
from aztts import AzTTS

tts = AzTTS(snapshot_download("ilqarrrr/Inflect_Micro_v2_Azerbaijan"))
tts.save("Salam, necəsiniz?", "out/salam.wav")

Normalisation and chunking run automatically. Calling the model/ package directly means calling normalize_az yourself.

Without PyTorch

onnx/ holds the same model as two graphs (the v2.1 ONNX export matched PyTorch to a parity error of 7e-9):

Graph Inputs Outputs
onnx/duration.onnx tokens, lengths, length_scale m_p_exp, logs_p_exp, y_mask
onnx/decode.onnx m_p_exp, logs_p_exp, y_mask, zp_noise, noise_scale waveform

This is what the browser playground runs, through ONNX Runtime Web. On the CPU it swaps decode.onnx for an int8 copy, built by tools/quantize_decoder.py in the GitHub repository. web/ in the GitHub repository is a working implementation, including the Azerbaijani text layer and the stress frontend ported to JavaScript.

Model details

Architecture VITS (compact), 9,356,513 parameters
Audio 24 kHz, mono, single speaker
Frontend eSpeak NG (az) through phonemizer, then the stress frontend (frontend_hook.py)
Starting point the first Azerbaijani checkpoint, all 410 tensors
Training main run to step 20k, then a decoder polish run; step 30,000 exported
Data 23,088 training clips distilled from a teacher TTS (471 more held out)
Speed 2-4x real time on CPU; ONNX real-time factor 0.25 on one thread

The text layer ahead of the model does three things the checkpoint cannot: bleeps obscenities, rewrites digits, Roman numerals, dates, units and abbreviations into spoken words, and cuts sentences into ~15-word chunks, because this model's intonation flattens towards the end of a long sentence. The stress frontend inside the package moves the stress marks eSpeak NG puts on nearly every word, with rules checked by a native speaker.

Measurements

On 471 validation sentences the model never trained on. "Teacher" is the original recording; 0.3 is the default variation.

Measure Teacher v1 v2.1, variation 0.667 v2.1, variation 0.3
Whisper word error, 416 number-free sentences 2.07% 4.84% 3.65% ~2.65% (0.333)
Word error, 1-5 word sentences 0.8% 9.2% 5.7% 3.2%
SQUIM PESQ 4.20 3.47 4.18 --
Harmonic clarity 2-4 kHz (buzz score, higher is better) 10.6 6.4 8.8 8.9
Fixed ring, share of frames 13% 20% (6 kHz) 15% 15%
Pitch range (semitones) 16.8 13.0 15.3 15.0
Yes/no question ending (semitones) +2.9 -3.0 -1.2 -1.8

Limitations

  • One voice. No voice cloning, no multi-speaker support, no emotion control.
  • At 9.36M parameters the speech is clear and intelligible but not fully natural, and intonation flattens over long sentences.
  • Yes/no questions end with falling pitch, where the teacher's rise. Only 677 training clips are such questions, and the model has no explicit pitch predictor.
  • m->n, b->d and q->k confusions remain, most often on a sentence's first and last word.
  • Corrected stress is carried by length, not yet by pitch.
  • The fixed 6 kHz ring of the first checkpoint is gone, but harmonic clarity at 2-4 kHz is still below the teacher's.
  • Rare names and unusual spellings are at the mercy of the phoneme frontend.
  • The training audio is synthetic, distilled from a teacher TTS, so the quality ceiling is the teacher. docs/MODEL.md in the repository has the details.

Licence and commercial use

Apache-2.0 for this project's code and for these weights, inherited from the base model. Commercial use is allowed, with one condition that comes from the phonemiser rather than the model.

Phonemisation calls phonemizer and eSpeak NG in-process, and both are GPL-3.0-or-later. Running them changes nothing. Shipping a combined work — a desktop app, a container image, a binary handed to customers — brings GPL-3.0 obligations for those components.

The checkpoint itself does not need eSpeak: config.json declares accepts_prephonemized_input: true, so phonemes can come from any frontend. THIRD_PARTY_NOTICES.md in the GitHub repository sets out both routes.

Training data. The first Azerbaijani checkpoint, which v2.1 starts from, was trained on ughurabbasov/azerbaijani-tts-dataset. It declares no licence file; asked in its Hugging Face community tab, the author confirmed free use and asked to be credited. Attribution is the condition of use — if you build on this model, credit the dataset too.

Files. The PyTorch export sits at the root (model.pth, config.json, the frontend and the VITS runtime); checksums.sha256 verifies 25 of them. NOTICES.md is this project's full third-party notice, while LICENSE and THIRD_PARTY_NOTICES.md are the upstream package's own, unchanged.

Ethical use

Do not use this voice to impersonate a real person or to produce deceptive content. Disclose synthetic speech where the context could otherwise mislead.

Credits and citation

Built on two openly published works: owensong/Inflect-Micro-v2 by Owen Song (the checkpoint) and ughurabbasov/azerbaijani-tts-dataset by ughurabbasov (the audio the first checkpoint was trained on). Please cite all three:

@software{azerbaijani_tts,
  title  = {Azerbaijani TTS: an offline 9.36M-parameter VITS model},
  author = {Huseynli, Ilqar},
  year   = {2026},
  url    = {https://github.com/HuseynliIlqar/Inflect_Micro_v2_Azerbaijan},
  note   = {Adapted from owensong/Inflect-Micro-v2 (Apache-2.0);
            first trained on ughurabbasov/azerbaijani-tts-dataset,
            used with the author's permission; v2.1 continued on
            audio distilled from a teacher TTS}
}
Downloads last month
128
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ilqarrrr/Inflect_Micro_v2_Azerbaijan

Quantized
(6)
this model

Dataset used to train ilqarrrr/Inflect_Micro_v2_Azerbaijan