Kothika-TTS

Model Description

Kothika-TTS is an Assamese focused text-to-speech model built on Zyphra's speech synthesis architecture. It converts Assamese Unicode text directly into natural sounding speech without transliteration or external text normalization.

The model supports fully local inference, zero shot voice cloning from reference audio, and 16 built in voices.

While Assamese is the primary focus, it also supports English, Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, and Urdu.

👉 Try it live: Interactive demo on Hugging Face Spaces

Getting started

Everything needed is inside this repository, clone and ready to use :

git clone https://huggingface.co/ananddey/kothika-tts.git
cd kothika-tts
pip install torch torchaudio numpy soundfile huggingface_hub tokenizers einops pyloudnorm descript-audio-codec==1.0.0 zyphra-textnorm==0.1.0

First synthesis downloads approx 8 GB of codec weights one time only, which is cached locally. CPU works but expect ~20–30 s generation time, a GPU works much faster.

from kothika_tts import KothikaTTS

tts = KothikaTTS(device="auto")                   # device="auto" (cuda if present, else cpu)

# built in voices (Anamika)
tts.speak("অসমীয়া ভাষাত কথা পতাটো সহজ নহয়।", voice="Anamika", out="out.wav")
tts.speak("নমস্কাৰ।", voice="Arup", out="namaskar.wav")

# reference voice (voice cloning)
tts.speak("অসম এখন প্ৰাকৃতিক সৌন্দৰ্যৰ ৰাজ্য।",
          voice="my_voice.wav", out="clone.wav",
          loudness=-18, fmax=16000, speaking_rate=35, temperature=0.8)

# conditioned delivery (dark & slow)
tts.speak("জ্ঞান আৰু শিক্ষাই মানুহক সহায় কৰে।", voice="Rajshri", out="dark.wav",
          loudness=-36, fmax=2500, speaking_rate=14, temperature=0.8)

Voices

16 named voices ship with this release, all verified by listening for clarity and Assamese pronunciation.

Female: Anamika · Bidisha · Sonali · Rajshri · Indrani · Anita · Jonali · Trishna · Isha

Male: Arup · Gautam · Joydip · Sameer · Diganta · Pranjal · Bishnu

Voice conditioning is stable across all sixteen, you choose who speaks.

Expressions

loudness — -36 dark & quiet, -18 bright & hot
fmax — 2500 mellow, 16000 crisp
speaking_rate — 15 slow, 35 brisk
temperature — 0.6 calm & steady, 1.0 natural

"condition_values": {"loudness": -18, "fmax": 16000, "speaking_rate": 30},
"temperature": 0.6

Leave blank for a natural default.

Demo audio

Assamese sentences, with expression pairings.

Demo	Voice · loudness/fmax/rate · temp	Text
1	Anamika · −24 · 12000 · 25 · 1.0	নমস্কাৰ, আপোনালোক সকলোকে মোৰ আন্তৰিক শুভেচ্ছা জনাইছোঁ।
2	Sameer · −18 · 16000 · 32 · 0.8	অসম এখন প্ৰাকৃতিক সৌন্দৰ্য, চহকী সংস্কৃতি আৰু বৈচিত্ৰ্যপূৰ্ণ পৰম্পৰাৰে ভৰপূৰ ৰাজ্য।
3	Jonali · −30 · 8000 · 16 · 1.0	আজি পুৱাৰ কোমল ৰ'দে চৌদিশ উজলাই তুলিছে, আৰু চৰাইৰ মাতে পৰিৱেশটো অধিক মনোমোহা কৰি তুলিছে।
4	Joydip · −18 · 16000 · 35 · 1.0	প্ৰযুক্তিৰ দ্ৰুত উন্নতিৰ লগে লগে কৃত্ৰিম বুদ্ধিমত্তাই আমাৰ জীৱন, শিক্ষা আৰু যোগাযোগৰ ক্ষেত্ৰত নতুন সম্ভাৱনাৰ সৃষ্টি কৰিছে।

Citation

If you use this model, please cite Kothika-TTS:

@misc{kothika_tts,
  title      = {{Kothika-TTS}},
  author     = {Anand Dey},
  year       = {2026},
  publisher  = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ananddey/kothika-tts}},
}

Downloads last month: -

ananddey
/

kothika-tts