Voxi Lite 1.0.0
CPU text-to-speech for Indian phone-agent calls, with streaming. Voxi Lite is built on
Mithi v1 Flash (Apache-2.0) with our own fine-tuned prosody writers
and voices; see NOTICE. Status: release candidate. See Before production below.
- Languages:
- Hindi, Tamil and English;
- Hinglish and Tanglish (romanised Hindi or Tamil mixed with English);
- English in Indian, American and British accents.
- Voices: 11. Seven are Indian (payal, arjun, varun, vijay, vysh, sanjana, vishnu), with 6 speaking styles each for payal and arjun; four are American and British English (harper, ethan, amelia, oliver).
- Sounds: laughs, "hmm", "umm", "mm-hmm", sighs and coughs in each Indian voice, from tags in the text
(
[laughs]) or plain fillers ("Hmm, ..."). - Size:
- Indian voices: 99 M parameters per voice. That is a 59 M decoder, a 21 M word encoder and a 19 M prosody writer; arjun has his own 19 M writer, and varun, vijay and vysh share another.
- American and British voices: Kokoro-82M, 82 M parameters.
- Full download: about 920 MB, including a 214 MB ONNX copy of the decoder for the fast CPU path.
- Weights: safetensors, bf16 for the decoder, word encoder and voice packs and fp32 for the prosody writers; Kokoro-82M is an fp32 PyTorch checkpoint.
- Output: 24 kHz mono audio, or 8 and 16 kHz, as WAV, PCM16 or G.711 μ-law.
- Runs on: CPU only, no GPU needed. On a laptop with 8 threads, it computes about 2× faster than real time.
- Licence: Apache-2.0, free to use, including commercially. See Licence.
Voices
| Voice | |
|---|---|
payal |
Female, warm phone-agent delivery. |
arjun |
Male, clear and professional. |
payal@<style>, arjun@<style> |
Speaking styles: happy, sad, angry, fearful, surprised, disgusted. Add -soft (half), -x2 or -x3 (stronger). For calls: @happy (warm) and @sad-soft (apologetic). |
varun, vijay |
Male. |
vysh |
Female. |
sanjana, vishnu |
Female and male. |
harper, ethan |
American English, female and male (lang: en only). |
amelia, oliver |
British English, female and male (lang: en only). |
All voices are synthetic: payal, arjun, varun, vijay and vysh were learned from Maya 2 Native's built-in personas, and the American and British voices are Kokoro-82M voices (Apache-2.0). Tell callers they are hearing a synthetic voice. The American and British voices have no speaking styles, and very short lines (two or three words) can sound weak; Hindi or Tamil text with them is rejected with a 400 error.
Languages:
hi,taanden;hinglish: romanised Hindi plus English;tanglish: romanised Tamil plus English.
Sounds and fillers
Write a tag in square brackets, or the filler itself, and Voxi plays a short recording of that sound in the same voice instead of reading it as a word.
| Write | Plays |
|---|---|
[laughs], [laugh], [chuckles], [giggles] |
a short laugh |
[laughs hard], [big laugh] |
a bigger laugh |
Hmm, [hmm], हम्म |
thinking "hmm" |
Mm-hmm, Uh-huh, [mhm] |
agreeing "mm-hmm" |
Umm, Uh, Erm, उम्म |
hesitation "umm" |
[sighs] |
a sigh |
[coughs], [clears throat] |
a cough |
Hmm, let me check that for you. [laughs] Okay, your booking for tomorrow at 10 AM is confirmed.
Each sound has several takes that rotate, so repeats do not sound identical. Unknown tags are dropped silently. Breaths are not available. The English-only voices (harper, ethan, amelia, oliver) have no sound recordings.
Emotions
Use an emotion as the voice ("voice": "payal@sorry") or write it in the text before the part it covers:
Hello, this is Payal. [sorry] I'm really sorry for the delay. [cheerful] Good news, your refund is approved!
| Emotion | Plays (payal, arjun) |
|---|---|
warm |
@happy-soft |
cheerful |
@happy |
sorry |
@sad-soft |
firm |
@angry-soft |
calm, neutral |
the plain voice |
happy, sad, angry, fearful, surprised, disgusted |
that style |
A tag lasts until the next tag. Voices without styles (sanjana, vishnu, varun, vijay, vysh, the English voices) read tagged parts in their plain voice.
Listen
Every clip was made by this release on a CPU, with the default settings (24 kHz).
| Hindi | Hinglish | Tamil | Tanglish | English | |
|---|---|---|---|---|---|
| Payal (female) | |||||
| Arjun (male) | |||||
| Sanjana (female) | |||||
| Vishnu (male) |
Lines: Hindi: नमस्ते, मैं Voxi से बोल रही हूँ। आपका बकाया ₹1,250 है, जो 15 तारीख़ तक भरना है। · Hinglish: Aapka order kal shaam tak deliver ho jayega, aur tracking link SMS par bhej diya hai. · Tamil: வணக்கம்! உங்கள் ஆர்டர் நாளை மாலைக்குள் வந்துவிடும். வேறு ஏதாவது உதவி வேண்டுமா? · Tanglish: Vanakkam! Unga order naalaikku varum, tracking link SMS-la anuppirukom. · English: Thank you for calling. Your refund of rupees 2,400 has been processed and will reach you in three to five days.
Launch clips (1.0)
| Clip | Text | Listen |
|---|---|---|
| payal · en | Hello, I'm Payal from Voxi. [cheerful] I can talk to your customers in Hindi, Tamil and English, right from a regular computer. [laughs] No GPU needed! | |
| payal · en | [sorry] I'm so sorry about the delay with your order. Hmm, let me check that for you. [cheerful] Good news! It's out for delivery, and it will reach you by 6 PM today. | |
| payal · hi | [warm] नमस्ते! मैं पायल बोल रही हूँ। आपका फिर से स्वागत है। बताइए, मैं आपकी क्या मदद कर सकती हूँ? | |
| payal · hinglish | Haan ji, aapka EMI ₹4,999 har mahine ki 5 tareekh ko auto-debit hoga. [laughs] Aur koi sawaal ho toh bataiye! | |
| arjun · en | [firm] This is a final reminder. Your payment of ₹12,500 is due by tomorrow evening. Please pay today to avoid late charges. | |
| arjun · hi | [cheerful] बधाई हो! आपका लोन मंज़ूर हो गया है। राशि कल तक आपके खाते में आ जाएगी। | |
| arjun · ta | வணக்கம்! உங்கள் ஆர்டர் நாளை காலை 11 மணிக்குள் டெலிவரி செய்யப்படும். நன்றி! | |
| varun · en | Mm-hmm, got it. Your booking for tomorrow at 10 AM is confirmed. See you then! | |
| vysh · hi | जी बिल्कुल। आपका OTP 482913 है। कृपया इसे किसी के साथ साझा न करें। | |
| vijay · tanglish | Sir, unga payment already process aayiduchu. Refund amount ₹1,500 next week-kulla account-la credit aagum. | |
| sanjana · ta | மன்னிக்கவும், உங்கள் கட்டணம் தோல்வியடைந்தது. தயவுசெய்து மீண்டும் முயற்சிக்கவும். | |
| harper · en | Thanks for calling! Your appointment is confirmed for tomorrow at 3 PM. | |
| ethan · en | Your package has shipped, and it should arrive on Friday. | |
| amelia · en | Lovely to hear from you. Your refund has been processed. | |
| oliver · en | Good afternoon. Your account has been updated successfully. |
English accents
American and British English voices (lang: en only).
| Agent line | Apology | |
|---|---|---|
| Harper (American, female) | ||
| Ethan (American, male) | ||
| Amelia (British, female) | ||
| Oliver (British, male) |
Lines: Thank you for calling. Your order shipped yesterday and should arrive by Thursday. Your tracking code is 482913. · I'm so sorry about the delay. I've issued a full refund, and you'll get a confirmation email within the hour.
Speaking styles
| Hindi | Tamil | English | |
|---|---|---|---|
Payal · warm (@happy) |
|||
Payal · apologetic (@sad-soft) |
|||
Arjun · warm (@happy) |
|||
Arjun · apologetic (@sad-soft) |
Lines: warm, Hindi: बधाई हो! आपका लोन अप्रूव हो गया है, अगले हफ़्ते तक पैसे आपके खाते में आ जाएँगे। · warm, Tamil: வாழ்த்துக்கள்! உங்கள் கடன் அங்கீகரிக்கப்பட்டது, அடுத்த வாரம் பணம் உங்கள் கணக்கில் வந்துவிடும். · warm, English: Congratulations! Your loan has been approved, and the money will reach your account next week.
Lines: apologetic, Hindi: माफ़ कीजिए, आपको जो परेशानी हुई उसके लिए हमें सच में खेद है। · apologetic, Tamil: மன்னிக்கவும், உங்களுக்கு ஏற்பட்ட சிரமத்திற்கு நாங்கள் உண்மையிலேயே வருந்துகிறோம். · apologetic, English: I am so sorry for the trouble you have had, and we truly regret it.
Numbers, read like an agent
| Line → how it is read | ||
|---|---|---|
| Hindi (Payal) | आपका OTP 482913 है। किसी भी मदद के लिए 1800 123 4567 पर कॉल करें। आपकी अगली किस्त ₹2.5 लाख की है, 10:30 बजे तक। → आपका OTP चार आठ दो, नौ एक तीन है। किसी भी मदद के लिए एक आठ शून्य शून्य, एक दो तीन, चार पाँच छह सात पर कॉल करें। आपकी अगली किस्त ढाई लाख रुपये की है, साढे दस बजे तक। |
|
| Tamil (Payal) | உங்கள் OTP 482913. உதவிக்கு 98765 43210 என்ற எண்ணை அழைக்கவும். உங்கள் அடுத்த தவணை ₹2,500. → உங்கள் OTP நான்கு எட்டு இரண்டு, ஒன்பது ஒன்று மூன்று. உதவிக்கு ஒன்பது எட்டு ஏழு ஆறு ஐந்து, நான்கு மூன்று இரண்டு ஒன்று பூஜ்ஜியம் என்ற எண்ணை அழைக்கவும். உங்கள் அடுத்த தவணை இரண்டாயிரத்து ஐந்நூறு ரூபாய். |
|
| English (Arjun) | Your OTP is 482913. For help, call 1800 123 4567. Your next instalment of Rs 2,500 is due at 10:30 on 15/12/2026. → Your OTP is four eight two, nine one three. For help, call one eight zero zero, one two three, four five six seven. Your next instalment of two thousand five hundred rupees is due at ten thirty on fifteen December two thousand twenty six. |
On a phone line
The lines above as 8 kHz audio, the way a caller hears them (format: mulaw, sample_rate: 8000 on the API gives the same signal, μ-law encoded):
| Hindi | Tamil | English |
|---|---|---|
What it does
- Numbers read like an agent.
- English and Tamil: phone numbers, OTPs and IDs in short digit groups.
- Hindi amounts: ₹1,250 → "बारह सौ पचास रुपये"; lakh and crore.
- Hindi times: साढ़े / सवा / पौने.
- Phone numbers, OTPs and IDs in short digit groups (Hindi and Tamil).
- Hinglish and Tanglish. Romanised Hindi or Tamil is written in its own script and English stays English.
- Word lists: 23k Hindi and 41k Tamil words, plus spoken forms common on calls.
- Unlisted words are approximated by rule (about 1 word in 10 in unseen Tanglish text).
- Streaming. A short first chunk for fast first audio; later chunks are sized from the machine's measured speed
so playback does not run dry. Fixed phrases can be pre-rendered (
warm) and start in under 1 ms. - Safe output.
- Silent, clipped or wrong-length audio is re-rendered automatically.
- Bad input gets a clear error.
- Emoji and control characters are dropped.
- Telephony formats. WAV, raw PCM16 or G.711 μ-law, at 8, 16 or 24 kHz.
- Sentence endings. Each sentence is read on its own, so every full stop ends like a sentence, with a 0.4 s pause
before the next (
VOXI_SENT_PAUSE). - Indian English. Words that American English and Indian English say differently follow the Indian form ("tomorrow" as tu-MORE-o), and times are read as letters ("10 AM" as "ten A M").
How it compares with other CPU models
The test set is 140 phone-agent lines: Hindi 40, Tamil 40, Hinglish 40 and Indian English 20. Each model ran on the same laptop CPU with 8 threads and was tested only in the languages it supports. Each line was rendered twice:
- written out: numbers already spelled as words, which tests clarity alone;
- as sent: the raw agent text ("₹1,250", phone numbers, IDs), which also tests the model's own number handling.
The table shows Whisper large-v3 character error rate (lower is better) as written out / as sent.
| Model | Hindi | Tamil | Indian English | Hinglish | CPU s per s of audio | Licence |
|---|---|---|---|---|---|---|
| Voxi Lite (payal) | 0.089 / 0.089 | 0.131 / 0.143 | 0.037 / 0.023 | 0.164 / 0.176 | 0.43 | Apache-2.0 |
| Mithi v1 Flash (charvi) | 0.094 / 0.098 | 0.150 / 0.173 | 0.043 / 0.029 | 0.294 / 0.264 | 0.57 | Apache-2.0 |
| goonj-1-82M | 0.072 / 0.061 | – | 0.041 / 0.029 | 0.182 / 0.177 | 0.37 | Apache-2.0 |
| Kokoro-82M (hf_alpha) | 0.084 / 0.079 | – | – | – | 0.39 | Apache-2.0 |
| Piper (hi_IN priyamvada) | 0.203 / 0.097 | – | – | – | 0.04 | voice card |
| Meta MMS-TTS | 0.171 / 0.260 | 0.114 / 0.202 | – | – | 0.29 | CC-BY-NC-4.0 |
A dash means the model has no voice for that language, or it was not run. Badly garbled lines (error rate above 0.5) on the as-sent text: Voxi Lite 2 of 140, Mithi 4 of 140, MMS-TTS 8 of 80.
- Where Voxi Lite leads:
- Tamil, Hinglish and Indian English on real agent text;
- the fewest garbled lines;
- the only model here that covers all five: Hindi, Tamil, Hinglish, Tanglish and Indian English.
- Where others lead:
- goonj, and to a lesser degree Kokoro, are clearer in Hindi.
- MMS-TTS is clearer in Tamil when numbers are already spelled out, but it drops digits in raw text and its licence is non-commercial.
- Piper is about 10× faster, but it has Hindi only.
- How lines were scored:
- Each line is matched against whichever written form the transcript is closer to, because Whisper writes a correctly spoken "twelve hundred fifty" as "1250".
- Hinglish is transcribed in Hindi mode, converted to Roman script with a standard scheme (IAST), and compared using one loose spelling rule, because in English mode Whisper translates Hindi speech instead of transcribing it. Treat the Hinglish scores as rough.
- What this does not show: naturalness. This table measures clarity only; a listening test is needed for that.
Run the server
docker build -t voxi-lite:1.0.0 .
docker run --rm -p 8080:8080 -e VOXI_THREADS=4 -e VOXI_API_KEY=change-me voxi-lite:1.0.0
# or without Docker (Python 3.11):
pip install --extra-index-url https://download.pytorch.org/whl/cpu -r requirements.txt
VOXI_WARM=warm_phrases.json uvicorn server:app --host 0.0.0.0 --port 8080
Each process or container holds one model, about 1.5 GB of RAM. Synthesis inside a process is serialised; scale with more processes.
| Setting | Default | Meaning |
|---|---|---|
VOXI_THREADS |
4 |
CPU threads per process |
VOXI_WARM |
(none) | JSON file of phrases per language, pre-rendered at start-up |
VOXI_WARM_VOICES |
payal,arjun |
Voices to pre-render |
VOXI_MODEL_DIR |
this folder | Model folder |
VOXI_API_KEY |
(none) | If set, every endpoint except /health and /ui needs it: Authorization: Bearer KEY or X-API-Key: KEY (WebSocket: ?key=KEY also works) |
VOXI_SENT_PAUSE |
0.45 |
Minimum silence after a full stop, in seconds |
VOXI_MKLDNN |
1 |
Set 0 if PyTorch crashes with a floating-point error on newer AMD EPYC CPUs |
VOXI_FIXES |
all on | Comma list of front-end and prosody fixes; "" turns them off for A/B tests |
API
# whole reply, WAV
curl -s localhost:8080/v1/tts -H 'content-type: application/json' \
-d '{"text": "आपका बिल ₹1,250 है।", "lang": "hi", "voice": "payal"}' -o reply.wav
# streamed 8 kHz mu-law for a phone line
curl -sN localhost:8080/v1/tts -H 'content-type: application/json' \
-d '{"text": "...", "lang": "hi", "voice": "arjun", "format": "mulaw", "sample_rate": 8000, "stream": true}' > reply.ulaw
# OpenAI-style
curl -s localhost:8080/v1/audio/speech -H 'content-type: application/json' \
-d '{"model": "voxi-lite", "input": "Your refund is processed.", "voice": "payal", "language": "en", "response_format": "wav"}' -o r.wav
| Endpoint | What |
|---|---|
GET /health |
{"status": "ok", "version", "voices"} |
GET /ui |
Browser test console: pick a voice and language, type text, hear it streamed with first-audio timing |
GET /v1/voices |
All voice names |
POST /v1/tts |
text, lang (hi | ta | en | hinglish | tanglish; default hi), voice (payal), format (wav | pcm16 | mulaw), sample_rate (8000 | 16000 | 24000), stream (false; streaming needs pcm16 or mulaw) |
POST /v1/audio/speech |
OpenAI shape: input, voice, response_format (wav | pcm = 24 kHz PCM16), plus language |
WS /v1/tts/ws |
Send {"text", "lang", "voice", "format", "sample_rate", "request_id"}. Receive binary audio frames, then {"event": "done", "request_id", "status", "audio_s", "first_audio_ms"}. Send {"event": "cancel"} to stop. Several requests per socket. |
Bad input (unknown voice, language or format; text over 2,000 characters) gets HTTP 400, or {"event": "error"}
on the WebSocket. Logs: one JSON line per request with lengths, language, voice, timings and status, never the text
or the audio.
Use from Python
from voxi import VoxiLite, encode
tts = VoxiLite() # this folder, 4 CPU threads
wav = tts.synthesize("आपका बिल ₹1,250 है।", lang="hi", voice="payal") # float32, 24 kHz
wav = tts.synthesize("Your refund has been processed.", lang="en", voice="amelia") # British English
tts.warm(["नमस्ते, मैं Voxi से बोल रही हूँ।"], lang="hi", voices=["payal", "arjun"])
for chunk in tts.stream("नमस्ते, मैं Voxi से बोल रही हूँ। आपका ऑर्डर कल पहुँचेगा।", lang="hi", voice="payal"):
phone.send(encode(chunk, "mulaw", 8000))
Tests
python tests/check_release.py # 22 edge cases + speed
python tests/test_server.py # API / WebSocket checks
python tests/fuzz.py --n 80 # random mixed input: no crashes, no bad audio
python tests/latency.py # first audio and playback gaps on THIS machine
Before production
- Native-speaker review of Hindi and Tamil (50 lines each): wrong words, stress, numbers.
- Measure on the production CPU with
tests/latency.py. Target: first audio under 0.5 s at p95 for short first sentences. Measured on an 8-vCPU cloud VM (AMD EPYC, 4 processes × 2 threads): first audio 0.28–0.36 s median, about 7× faster than real time, and 4 simultaneous replies at 0.55 s median first audio. A busy laptop measured 1.1 s. - Listening check on a real phone line with 8 kHz μ-law.
Known limits:
- Styles are mild: clearest in Hindi, weak in Tamil.
- Arjun's Tamil scores slightly below sanjana/vishnu by ASR.
- Long clauses without commas can leave a short gap on slow CPUs.
- varun, vijay and vysh have no speaking styles yet.
Licence
Voxi Lite is released under the Apache License 2.0 (see LICENSE).
- Use: personal, research and commercial projects alike.
- Freedoms: you can use, modify and redistribute the model and code, and build products on it.
- Conditions: keep the licence and the credits in
NOTICEwith any copy you share. - Voices: all are synthetic. Tell callers they are hearing a synthetic voice, and do not present a voice as a real person.
Built with (open-source models and data; full credits in NOTICE)
| Part | What it is used for | Licence |
|---|---|---|
| Mithi v1 Flash | Base model: decoder, word encoder and prosody writer (writer fine-tuned for Voxi) | Apache-2.0 |
| Kokoro-82M | Mithi's decoder design; the American and British voices | Apache-2.0 |
| indic-bert | Word encoder inside Mithi | MIT |
| Maya 2 Native | Generated the speech that the payal, arjun, varun, vijay and vysh voices were trained on | Apache-2.0 |
| CosyVoice 3 | Generated the laugh and filler recordings, cloning Voxi's own voices (not needed to run Voxi) | Apache-2.0 |
| AI4Bharat Rasa | Speaking styles and voice texture | CC-BY-4.0 |
| NVIDIA Nemotron-Indic | Hinglish and Tanglish word lists (CC-BY-4.0 rows only) | CC-BY-4.0 |
| goonj-1-82M | Hinglish word list | Apache-2.0 |
- Downloads last month
- 39
Model tree for IOTEverythin/voxi-lite
Base model
Oomni/mithi-v1-flash