Uzbek TTS · Chatterbox fine-tune

Uzbek text,
natural voice.

A text-to-speech model fine-tuned on ~10 hours of a single audiobook narrator. It splits long text into sentences on its own and keeps natural pauses and intonation. It can also borrow another voice from a ~10-second recording.

The audio in the main sample is raw model output, with no editing.

All samples Get in touch

Main sample

0:00—

WAV

The highlighted word follows playback (approximate). Click any word to jump there.

of data: ~10h
voices: 1
sample rate: 24kHz
samples: 6

01

All samples

From a greeting to a dialogue — texts in different tones. Click a row to open it in the player above.

Keyboard

Space K: play / pause
J L: back / forward 5 seconds
N P: next / previous sample
V: switch voice
← →: 1 second, when the waveform is focused
↑ ↓: move through the list

02

Results

Every number can be checked: the texts and audio are on this page, and the evaluation script is in the repo. The evaluation is still small — 6 sentences, 105 words — so treat these numbers as a direction, not a final benchmark.

Intelligibility — round-trip WER / CER

Each sample was transcribed back with an Uzbek ASR model (rustam1221/uzbek-asr-gigaam, 220M + KenLM) and compared with the original text. Both sides are lowercased, apostrophes are canonicalized and punctuation is stripped.

Voice	Words	WER	CER	Only error
Narrator	105	1.0%	0.1%	«esar» → «esa»
Female voice	105	1.0%	0.7%	sentence-final «edim» swallowed

Timbre similarity

Cosine similarity from Chatterbox's own voice encoder (1.00 = the same voice). The female voice is cloned without retraining, from a ~12-second recording.

Synthesis	vs narrator reference	vs female reference
Narrator	0.95	0.60
Female voice	0.62	0.94
The two references	0.63

The narrator's timbre does not leak into the female voice: its similarity to the narrator (0.62) is at the level of two different people (0.63).

Speaking rate and pitch

Audio	Rate, chars/s	F0 median	Pitch range
Narrator reference	—	102 Hz	—
Narrator synthesis	14.0	101 Hz	7.2 st
Female reference	14.7	251 Hz	—
Female synthesis	16.8	236 Hz	7.6 st

st — semitones, 10th–90th percentile range. The female voice speaks 14% faster than the original recording: the model learned the narrator's rhythm, and the cloned timbre speeds it up. This is the biggest open issue — slowing it down with cfg_weight 0.3 is being tested.

03

How it works

Five steps from text to sound. Only the text model (T3) was retrained for Uzbek; everything else is stock Chatterbox.

1
Normalization
Text is cleaned, numbers and abbreviations are spelled out, punctuation is normalized.
2
Chunking
Long text is split into short chunks at sentence boundaries.
3
T3 · text → tokens
A Llama-based model turns each chunk into speech tokens. The narrator's voice is learned here, with LoRA.
4
S3Gen · tokens → audio
A 24 kHz waveform is generated from the tokens; the timbre comes from a reference recording.
5
Post-processing
Extra silence and breath tails are trimmed, and chunks are joined with natural pauses.

04

Data

A ~12-hour audiobook read by one narrator was turned into a training set automatically — not a single transcript was written by hand.

clips: 5,005
hours: 10.07
words: 65,731
unique words: 16,954

Preparation

Segmentation. Silero VAD cuts the audio at sentence boundaries into 1.5–12 s clips, 24 kHz.
Transcription. Each clip is transcribed, with punctuation, by an Uzbek ASR model (GigaAM 220M + KenLM).
Filtering. Ads, clipped audio and clips with an implausible speaking rate are dropped.
Split. 4,955 clips for training; 50 held out for evaluation and never seen in training.

Rejected

Reason	Clips
Ads or misrecognized	42
Clipping	8
Implausible rate	5
Total (of 5,060)	55 · 1.1%

Clip length

0–2 s28
2–4 s441
4–6 s928
6–8 s1,429
8–10 s2,014
10–12 s165

Median 7.6 s, mean 7.2 s, about 13 words per clip.

Because the transcripts also come from ASR, they contain errors — for example, the book's «teleekran» was often heard as «Telegram». The model may learn those errors too; hand-checked transcripts are a next step.

05

Training

Only T3, the part that turns text into speech tokens, was retrained; the waveform generator (S3Gen) and the voice encoder are frozen upstream Chatterbox weights.

Settings

Base model: ResembleAI Chatterbox (T3 + S3Gen + VE)
Trained modules: T3 only; S3Gen and VE frozen
Method: LoRA r=128, α=256 on attention and MLP layers; text embedding and output head fully trained
Tokenizer: Uzbek, 2,454 tokens
Optimization: lr 1e-4, effective batch 32, fp16
Hardware: Google Colab T4 (16 GB)
Steps: 587 / 770 (5 epochs planned)
Checkpoint: step 400 (~2.6 epochs), LoRA merged into the weights

Loss

Step	Epoch	Loss	LR
154	1.0	6.15	8.0e-5
308	2.0	5.23	6.0e-5
400 · shipped	2.6	—	4.8e-5
587 · session ended	3.8	—	—

LR decays linearly from 1e-4 to step 770. Loss was logged at the end of each epoch; no separate validation set was used during training.

What we learned

10 hours is enough for one voice. After 400 LoRA steps, the ASR model understands the output with a 1% word error rate.
Free GPUs are a constraint. The Colab session ended at step 587/770, and with Drive full the latest saved copy was step 400. These results come from a model that has not converged.
The data was transcribed by ASR. The automatic route is cheap — only 55 of 5,060 clips were dropped — but ASR errors end up in the text too.
Timbre transfers, rhythm doesn't. A 12-second recording gives the voice colour with 0.94 similarity, but pace and stress stay the narrator's.

06

Synthesis speed

RTF is synthesis time divided by audio duration; above 1 means slower than real time. Measured while generating the 12 samples on this page.

Device	Audio	Synthesis	Mean RTF	Best
Apple M4, 16 GB (MPS)	95.9 s	736 s	7.7×	2.7×
CUDA GPU	not measured yet

On a Mac the result depends heavily on memory pressure: the model takes ~5 GB, and next to another heavy app RTF drops to 20×. The service is meant to run on a CUDA server in production.

07

Using the API

Synthesis runs through a queue: a request returns a job_id immediately, and every finished segment is downloaded as its own WAV. Write to us for an API key.

08

Files and versions

Layout

models/
├── t3_finetuned_merged.safetensors  2.0 GB  Uzbek T3, LoRA merged
├── s3gen.safetensors                1.0 GB  upstream Chatterbox
├── ve.safetensors                   5.4 MB  voice encoder
├── tokenizer.json                    72 KB  Uzbek, 2,454 tokens
├── conds.pt                         108 KB  default conditioning
└── voices/                          reference recordings

Weights are never downloaded at runtime — the service reads them from a local folder.

Versions

Version	Weights	Status
v0.1	step 400, 2.6 epochs	current · 2026-10-17
v0.2	full 770 steps	planned

09

Limitations

The model was trained on the narrator only; the female voice is cloned without retraining and speaks 14% faster.
A short sentence-final word is sometimes swallowed (1 of 12 samples).
Numbers followed by a suffix (e.g. a time plus «da») are not yet read correctly.
Training did not finish (587 / 770 steps; step 400 is shipped).
Evaluation covers only 6 sentences so far; there is no large benchmark yet.

10

Next steps

Round-trip WER on the 50 held-out sentences.
Train to step 770 and compare.
Slow down the female voice (cfg_weight 0.3).
Fix normalization of numbers with suffixes.
Measure speed on a CUDA server.

11

Citation

BibTeX

@misc{uzbek_tts_narrator,
  title  = {Uzbek TTS: Chatterbox T3 LoRA fine-tune on a single narrator},
  author = {abbos2332},
  year   = {2026},
  url    = {https://huggingface.co/spaces/abbos2332/uzbek-tts-demo}
}

Contact

Want to work with the model?

Write about an API key, voicing your own text, or a partnership.

abbosfayzullayev202@gmail.com

Send an email

Uzbek text,natural voice.

All samples

Results

Intelligibility — round-trip WER / CER

Timbre similarity

Speaking rate and pitch

How it works

Normalization

Chunking

T3 · text → tokens

S3Gen · tokens → audio

Post-processing

Data

Preparation

Rejected

Clip length

Training

Settings

Loss

What we learned

Synthesis speed

Using the API

Files and versions

Layout

Versions

Limitations

Next steps

Citation

Want to work with the model?

Uzbek text,
natural voice.