Modelpublic

MukhammedTogmanov/Til-TTS

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

1likes73downloads

Model Card

Til-TTS

Til-TTS is a Kazakh text-to-speech model fine-tuned directly from k2-fsa/OmniVoice. The model synthesizes natural Kazakh speech and supports controllable voice attributes (gender, age, pitch) as well as expressive, contextual non-verbal speech reactions.

Overview

Til-TTS was trained on a large, quality-verified Kazakh speech corpus, combining professionally recorded studio data, large-scale broadcast and crowdsourced speech, and purpose-built synthetic data for voice control and expressive speech.

Data Pipeline

Final training corpus: 546,488 samples, assembled and verified through the following pipeline:

1.Source aggregation - combined five licensed Kazakh speech corpora: ISSAI KazakhTTS2, ISSAI KSC2 (six domains: crowdsourced, radio, parliament, podcasts, talkshow, and TV news), Google FLEURS, and Mozilla Common Voice.
2.Audio quality filtering - SNR and clipping-based filtering applied across all real-speech sources.
3.Deduplication - exact and near-duplicate transcripts capped to prevent overrepresentation of repeated phrases.
4.Automated ASR verification - every audio-text pair in the corpus was independently transcribed using GigaAM-Multilingual and compared against its reference transcript; pairs with high word error rate were excluded from training.
5.Text normalization and punctuation restoration - transcripts for the broadcast/crowdsourced portion of the corpus were upgraded from unpunctuated text to properly capitalized, punctuated text, improving the model's prosody and pacing.
6.Synthetic voice-control data - 18,000 samples spanning 30 controllable voice categories (2 genders x 5 age styles x 3 pitch levels), generated using the base model's voice-design interface.
7.Expressive speech data - a curated dataset of natural Kazakh speech containing contextual non-verbal reactions (laughter, sighs, dissatisfaction) was incorporated to teach the model to produce these reactions appropriately within continuous speech.

Training Data Breakdown

Source	Samples
ISSAI KSC2 - crowdsourced	264,114
ISSAI KSC2 - TV news	77,346
ISSAI KazakhTTS2	69,885
ISSAI KSC2 - talk shows	30,922
ISSAI KSC2 - podcasts	28,891
ISSAI KSC2 - parliament	22,232
Synthetic voice-control data	18,000
ISSAI KSC2 - radio	17,330
Expressive speech data	14,421
FLEURS	1,863
Common Voice	1,484

Training Configuration

—Base model: k2-fsa/OmniVoice
—Method: full fine-tuning (all parameters updated)
—Steps: 20,000
—Learning rate: 3e-5, cosine schedule
—Precision: bf16

Evaluation

Evaluated on a 50-sentence Kazakh test set spanning everyday speech, technical terminology, numbers, place names, and questions, using an independent ASR system for objective scoring.

Metric	Value
WER	0.0703
CER	0.0474
UTMOS (naturalness, 1-5 scale)	3.02

Voice-design control (gender, age, pitch) and expressive speech reactions were verified through direct listening evaluation.

Controllable Voice Attributes

Supported via the instruct parameter:

—Gender: male, female
—Age style: child, teenager, young adult, middle-aged, elderly
—Pitch style: low pitch, moderate pitch, high pitch

Expressive Speech

Til-TTS can produce contextual non-verbal reactions inline within generated speech, using bracketed tags in the input text:

[laughter], [sigh], [confirmation-en], [question-en], [surprise-ah], [surprise-oh], [dissatisfaction-hnn], and related tags.

Usage

bash

pip install omnivoice soundfile torch

python

from omnivoice import OmniVoice
import soundfile as sf
import torch

model = OmniVoice.from_pretrained(
    "MukhammedTogmanov/Til-TTS",
    device_map="cuda:0",
    dtype=torch.float16,
)

audio = model.generate(
    text="Бүгін біз жаңа қазақша дауыс моделін тексеріп жатырмыз.",
    language="Kazakh",
)
sf.write("output.wav", audio[0], 24000)

audio = model.generate(
    text="Бұл қазақша мәтінді дыбыстау мысалы.",
    language="Kazakh",
    instruct="female, teenager, high pitch",
)
sf.write("voice_design.wav", audio[0], 24000)

audio = model.generate(
    text="Керемет, [laughter] біз келісімге келдік ақыры.",
    language="Kazakh",
)
sf.write("expressive.wav", audio[0], 24000)

Limitations

—Very short utterances (one to two words) may synthesize with reduced reliability.
—Whole-utterance discrete emotion control (e.g., a single "happy" or "sad" mode) is not supported; expressiveness is achieved through contextual inline reactions rather than a global emotion parameter.
—As with any model trained on broadcast and crowdsourced audio alongside studio recordings, output acoustic character may vary somewhat depending on the style of the input text.

Audio Examples

Female teenager

Text: Мен қазақша дауыс моделін тексеріп жатырмын.

Female middle-aged

Text: Отбасыммен бірге демалыс күндерін өткіземіз.

Male teenager

Text: Бүгін достарыммен кездесуге барамын.

Male child (high pitch)

Text: Мен ойыншықтармен ойнағанды жақсы көремін.

Male elderly (low pitch)

Text: Өмірде көп нәрсені көрдім, көп тәжірибе жинадым.

Male, high pitch

Text: Бұл қазақша мәтінді дыбыстау мысалы.

Male, low pitch

Text: Бұл қазақша мәтінді дыбыстау мысалы.

Expressive - Laughter

Text: Керемет, біз келісімге келдік ақыры.

Expressive - Sigh

Text: Ет жүрегім езілді қан жылады

Expressive - Surprise

Text: Апырау, мұны күтпеген едім!

Expressive - Question tone

Text: Сіз бұл туралы не ойлайсыз?

Expressive - Confirmation

Text: Иә, түсінікті, солай жасаймыз.

Contact

it.technologier@gmail.com