Mithi v1 Flash

A small text-to-speech model for 11 Indian languages, with two voices: Satvik (male) and Charvi (female).

Languages: Hindi, English, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu.
Size: 99.1 M parameters, 197 MB.
Weights: bf16 safetensors.
Output: 24 kHz mono audio.

Quick start

1. Download the model and install the dependencies (Python 3.10 or newer):

pip install -U huggingface_hub
hf download Oomni/mithi-v1-flash --local-dir mithi-v1-flash
cd mithi-v1-flash
pip install -r requirements.txt
python -m spacy download en_core_web_sm

espeak-ng comes with the espeakng-loader package, so no system install is needed. A GPU is used if there is one; the model also runs on CPU.

2. Run it from the command line:

python inference.py --lang hi --voice charvi --text "आज मौसम कितना अच्छा है! चलो, शाम को कहीं बाहर घूमने चलते हैं।" --out hindi.wav
python inference.py --lang en --voice satvik --text "Your appointment is confirmed for Monday at 10:30 AM." --out english.wav

3. Or from Python, inside the mithi-v1-flash folder:

from mithi_tts import Mithi

tts = Mithi(".")                                   # the downloaded model folder
audio = tts.synthesize("नमस्ते! आप कैसे हैं?", lang="hi", voice="satvik")
tts.save(audio, "namaste.wav")                     # 24 kHz WAV

for lang, text in [("ta", "வணக்கம்! எப்படி இருக்கீங்க?"), ("bn", "নমস্কার! কেমন আছেন?")]:
    tts.save(tts.synthesize(text, lang=lang, voice="charvi"), f"hello_{lang}.wav")

4. Run on CPU (laptops, Raspberry Pi, phones, servers without a GPU):

python inference.py --device cpu --threads 4 --lang hi --voice charvi --text "आज मौसम कितना अच्छा है! चलो, शाम को कहीं बाहर घूमने चलते हैं।" --out hindi_cpu.wav

--threads sets how many CPU cores to use; without it PyTorch uses all of them. From Python:

import torch
torch.set_num_threads(4)

from mithi_tts import Mithi
tts = Mithi(".", device="cpu")
tts.save(tts.synthesize("नमस्ते! आप कैसे हैं?", lang="hi", voice="charvi"), "namaste_cpu.wav")

Measured speed for an average 15-word sentence (about 7 s of speech), on a server CPU (AMD EPYC 9554):

CPU threads	Time per sentence
1	~8 s
4	~3 s

Slower CPUs, such as a Raspberry Pi, take proportionally longer.

Languages: hi Hindi · en English · bn Bengali · gu Gujarati · kn Kannada · ml Malayalam · mr Marathi · or Odia · pa Punjabi · ta Tamil · te Telugu

Voices: satvik (male) and charvi (female). Long text is read sentence by sentence automatically.

Voices

One male and one female speaker. Their training speech was generated with Maya2-native inference.

Voice	Speaker
Satvik	Male
Charvi	Female

Samples

Every sentence in both voices: 11 languages, one conversational line and one news or education line with numbers. The text is shown as typed; numbers, times and currency are expanded into words by the text normaliser.

Conversational

Sentence	Satvik (male)	Charvi (female)
Hindiअरे, तुम आ गए! बहुत दिनों बाद मिले हो। चलो, पहले एक कप चाय पीते हैं, फिर आराम से बातें करेंगे।
EnglishWait, you actually got the tickets? That's amazing! I can't believe we're finally going to see them perform live.
Bengaliমা, আজ রাতে কী রান্না হচ্ছে? গন্ধটা তো দারুণ লাগছে! আমি কিন্তু খুব খিদে নিয়ে বাড়ি ফিরেছি।
Gujaratiકેમ છો? ઘણા દિવસે મળ્યા! ચાલો, આજે સાંજે સાથે ચા પીવા જઈએ.
Kannadaನಮಸ್ಕಾರ! ನಿಮ್ಮ ಆರ್ಡರ್ ಈಗಾಗಲೇ ಹೊರಟಿದೆ, ನಾಳೆ ಮಧ್ಯಾಹ್ನದೊಳಗೆ ನಿಮ್ಮ ಮನೆಗೆ ತಲುಪುತ್ತದೆ. ಇನ್ನೇನಾದರೂ ಸಹಾಯ ಬೇಕಾ?
Malayalamഹേയ്, എത്ര നാളായി കണ്ടിട്ട്! സുഖമാണോ? വാ, ഇന്ന് വൈകുന്നേരം നമുക്ക് ഒരുമിച്ച് ചായ കുടിക്കാം.
Marathiआई, आज संध्याकाळी आपण सगळे मिळून बाहेर जेवायला जाऊया का? खूप दिवस झाले, सगळे एकत्र बसलो नाही.
Odiaଆରେ, କେତେ ଦିନ ପରେ ଦେଖା ହେଲା! କେମିତି ଅଛ? ଚାଲ, ଆଜି ସନ୍ଧ୍ୟାରେ ଏକାଠି ଚା ପିଇବା।
Punjabiਕੀ ਹਾਲ ਹੈ ਜੀ? ਬਹੁਤ ਦਿਨਾਂ ਬਾਅਦ ਮਿਲੇ ਹਾਂ! ਆਓ, ਅੱਜ ਸ਼ਾਮ ਨੂੰ ਇਕੱਠੇ ਚਾਹ ਪੀਂਦੇ ਹਾਂ।
Tamilஹேய், எவ்வளவு நாளாச்சு பார்த்து! எப்படி இருக்கீங்க? வாங்க, இன்னைக்கு சாயங்காலம் ஒன்னா காபி குடிக்கலாம்.
Teluguమీ టికెట్ కన్ఫర్మ్ అయింది! రైలు రేపు ఉదయం ఆరు గంటలకు బయలుదేరుతుంది, కాబట్టి కొంచెం ముందుగానే స్టేషన్‌కి రండి.

News and education, with numbers

Sentence	Satvik (male)	Charvi (female)
Hindiभारतीय अंतरिक्ष अनुसंधान संगठन ने आज सुबह 9:30 बजे श्रीहरिकोटा से एक नया उपग्रह सफलतापूर्वक लॉन्च किया, जिसकी लागत लगभग ₹450 करोड़ है।
English · EducationLight from the Sun takes about 8 minutes and 20 seconds to reach the Earth, travelling nearly 150 million kilometres.
Bengaliরাজ্যের মাধ্যমিক পরীক্ষায় এ বছর 92 শতাংশ ছাত্রছাত্রী পাশ করেছে, যা গত বছরের চেয়ে 3 শতাংশ বেশি।
Gujaratiગુજરાતમાં આ વર્ષે 1,250 નવી શાળાઓ શરૂ કરવામાં આવશે, જેનાથી આશરે 2 લાખ બાળકોને લાભ મળશે.
Kannadaಬೆಂಗಳೂರಿನ ಮೆಟ್ರೋದಲ್ಲಿ ಕಳೆದ ತಿಂಗಳು ಪ್ರತಿದಿನ ಸರಾಸರಿ 8 ಲಕ್ಷ ಜನರು ಪ್ರಯಾಣಿಸಿದ್ದಾರೆ ಎಂದು ಅಧಿಕಾರಿಗಳು ತಿಳಿಸಿದ್ದಾರೆ.
Malayalamകേരളത്തിലെ ഡിജിറ്റൽ സാക്ഷരതാ പദ്ധതിയിലൂടെ 65 വയസ്സിന് മുകളിലുള്ള 2 ലക്ഷത്തിലധികം പേർ സ്മാർട്ട്ഫോൺ ഉപയോഗിക്കാൻ പഠിച്ചു.
Marathiपुण्यात आज सकाळी 7 वाजता 12 अंश सेल्सिअस तापमानाची नोंद झाली, जे या हिवाळ्यातील सर्वात कमी तापमान आहे.
Odiaଓଡ଼ିଶାରେ ଏହି ବର୍ଷ 3,500 ନୂଆ ଶିକ୍ଷକ ନିଯୁକ୍ତ ହେବେ ବୋଲି ସରକାର ଆଜି ଘୋଷଣା କରିଛନ୍ତି।
Punjabiਪੰਜਾਬ ਵਿੱਚ ਇਸ ਸਾਲ ਕਣਕ ਦੀ ਪੈਦਾਵਾਰ 180 ਲੱਖ ਟਨ ਤੱਕ ਪਹੁੰਚ ਗਈ ਹੈ, ਜੋ ਪਿਛਲੇ ਸਾਲ ਨਾਲੋਂ 5 ਫ਼ੀਸਦੀ ਵੱਧ ਹੈ।
Tamilசென்னையில் நேற்று 45 மில்லிமீட்டர் மழை பதிவானது, மேலும் அடுத்த 2 நாட்களுக்கு மழை தொடரும் என்று வானிலை ஆய்வு மையம் தெரிவித்துள்ளது.
Teluguహైదరాబాద్‌లో ఈ ఏడాది 2,000 కొత్త ఎలక్ట్రిక్ బస్సులు ప్రవేశపెట్టనున్నట్లు ప్రభుత్వం ప్రకటించింది.

All clips and their text (samples/sentences.tsv) are in the samples/ folder.

Model

text
 └─ text normaliser               numbers, dates, currency, URLs ... → words
 └─ espeak-ng / misaki G2P        → phonemes
 └─ indic-bert                    one meaning vector per word
 └─ prosody writer                per phoneme: duration, voicing, pitch, loudness, breaths
 └─ Kokoro text encoder + iSTFTNet decoder → 24 kHz audio

Component	Parameters
indic-bert (ALBERT, word table trimmed to our languages)	21.3 M
Prosody writer (Transformer reader + autoregressive writer)	18.8 M
Kokoro text encoder + iSTFTNet decoder (fine-tuned)	58.9 M

The prosody writer reads the sentence and writes its prosody one phoneme at a time: how long each sound lasts, its pitch contour, its loudness, and where to breathe. The Kokoro decoder turns that prosody into audio in the chosen voice.

Inference settings

These are in config.json, field inference:

Setting	Value
Writer temperature, top-p	0.7, 1.0
Voicing	greedy (temperature 0)
Silence at each end	at most 50 ms
Pitch shift	+0.6 semitone
Random seed	0 for every sentence (same text, same audio)
Long text	read sentence by sentence, joined with 0.15 s of silence

Files

File	Contents
`inference.py`	command-line script: text in, WAV out
`mithi_tts/`	the Python package that runs the model
`requirements.txt`	Python dependencies
`config.json`	languages, voices, inference settings
`decoder.safetensors`, `kokoro_config.json`	Kokoro text encoder and iSTFTNet decoder
`writer.safetensors`	prosody writer
`voices.safetensors`	Satvik and Charvi
`indic_encoder/`	indic-bert
`samples/`	the example clips above and their text
`assets/`	banner and wordmark

Limitations

Less expressive than Maya2-native. Intonation and emotion are flatter than in the speech it learned from.
Kannada, Malayalam, Marathi, Odia: speech can come out a little faster than natural.
Pronunciation depends on espeak-ng / misaki: a word the G2P gets wrong is spoken wrong.
Numbers, dates, currency, URLs are read only as well as the text normaliser expands them.
Long text is read sentence by sentence; each sentence's prosody is written without seeing the others.
No published evaluation yet: no automated scores and no human listening study (MOS).

Acknowledgements

Maya2-native by Maya Research (Apache-2.0): the training speech for the Satvik and Charvi voices was generated with it.
Kokoro-82M text encoder and iSTFTNet decoder: hexgrad/kokoro, Apache-2.0.
indic-bert: ai4bharat/indic-bert, MIT.
G2P: misaki and espeak-ng.