Orqanta-Voice

Ultra-Fast Streaming Multilingual Text-to-Speech Engine for 11 Indic Languages & English

⚡ Executive Summary

Orqanta-Voice is an ultra-fast, production-grade streaming text-to-speech (TTS) engine purpose-built for 11 major Indic languages and Indian English.

Engineered with continuous-batching CUDA graphs, Orqanta-Voice achieves a state-of-the-art ~14 ms Time-to-First-Audio (TTFA) at batch size 1 on an H100 and serves 224 concurrent streams with 0 playback stalls. It pairs an autoregressive Qwen3 (0.6B) language backbone with a 24 kHz Mimi neural audio codec, a dedicated 8-layer Depth Transformer, and ECAPA-TDNN zero-shot speaker adaptation.

Whether you need conversational bots, interactive IVR systems, real-time voice assistants, or studio-grade narration, Orqanta-Voice provides natural human cadence, seamless zero-shot voice cloning (3–10s audio), and native text normalization for numbers, currency, and dates.

🎧 Audio Showcase & Multilingual Demos

Listen to real audio samples generated directly by Orqanta-Voice across different languages and voices:

Language	Voice & Character	Sample Transcript	Audio Preview & Download
🇬🇧 English	`Siddhant` (Adult Male, Professional)	"Welcome to the future of voice technology with Orqanta AI. Experience ultra-fast, real-time streaming speech across eleven Indic languages..."	▶️ Download WAV
🇮🇳 Odia (ଓଡ଼ିଆ)	`Vaidehi` (Adult Female, Gentle)	"ଓରକାନ୍ତା ଏଆଇ ତରଫରୁ ଆପଣ ସମସ୍ତଙ୍କୁ ସ୍ୱାଗତ। ଆମର ନୂତନ ଓରକାନ୍ତା ଭଏସ୍ ଇଞ୍ଜିନ୍ ସହିତ ଏବେ ୧୧ଟି ଭାରତୀୟ ଭାଷାରେ ଅତି ସ୍ପଷ୍ଟ କଣ୍ଠସ୍ୱର ଉପଲବ୍ଧ..."	▶️ Download WAV
🇮🇳 Odia (ବାତ୍ୟା ସତର୍କତା)	`Vaidehi` (Adult Female, News Alert)	"ବାତ୍ୟା ଅର୍ଣ୍ଣବ ଏବେ ଓଡ଼ିଶା ଉପକୂଳ ମୁହାଁ। ଆସନ୍ତା ୨୪ ଘଣ୍ଟା ମଧ୍ୟରେ ପବନର ବେଗ ୧୨୦ କିଲୋମିଟର ଏବଂ ରିଲିଫ୍ ପାଇଁ ₹୨୦୦ କୋଟି ଟଙ୍କା..."	▶️ Download WAV
🇮🇳 Hindi (हिन्दी)	`Vaidehi` (Adult Female, Expressive)	"नमस्कार! ओरकांता एआई की दुनिया में आपका स्वागत है। प्रस्तुत है ओरकांता वॉइस, भारत का सबसे तेज़ और आधुनिक वॉइस एआई इंजन..."	▶️ Download WAV
🇮🇳 Telugu (తెలుగు)	`Sreekar` (Adult Male, Conversational)	"నమస్కారం! ఓర్కాంటా ఏఐ కి స్వాగతం. భారతదేశపు అత్యుత్తమ రియల్-టైమ్ వాయిస్ టెక్నాలజీతో మీ ముందుకు వచ్చింది ఓర్కాంటా వాయిస్..."	▶️ Download WAV
🇮🇳 Tamil (தமிழ்)	`Kayal` (Adult Female, Clear)	"வணக்கம்! ஓர்காண்டா ஏஐ உங்களை அன்புடன் வரவேற்கிறது. இந்தியாவின் அதிவேக ஸ்ட்ரீமிங் வாய்ஸ் தொழில்நுட்பம் ஓர்காண்டா வாய்ஸ்..."	▶️ Download WAV
🇮🇳 Bengali (বাংলা)	`Anirban` (Adult Male, Steady Narrator)	"নমস্কার! ওরকান্তা এআই-এর জগতে আপনাকে স্বাগতম। উপস্থাপন করছি ওরকান্তা ভয়েস, ভারতের দ্রুততম মাল্টিলিঙ্গুয়াল স্পিচ ইঞ্জিন..."	▶️ Download WAV
🇮🇳 Gujarati (ગુજરાતી)	`Vaidehi` (Adult Female, Calm)	"નમસ્તે! ઓરકાન્તા એઆઈ માં આપનું હાર્દિક સ્વાગત છે. ઓરકાન્તા વૉઇસ સાથે હવે ૧૧ ભારતીય ભાષાઓમાં મેળવો સૌથી ઝડપી અને કુદરતી અવાજ..."	▶️ Download WAV
🇮🇳 Punjabi (ਪੰਜਾਬੀ)	`Nakul` (Young Male, Energetic)	"ਸਤਿ ਸ੍ਰੀ ਅਕਾਲ! ਓਰਕਾਂਤਾ ਏਆਈ ਵਿੱਚ ਤੁਹਾਡਾ ਨਿੱਘਾ ਸੁਆਗਤ ਹੈ। ਪੇਸ਼ ਹੈ ਓਰਕਾਂਤਾ ਵੌਇਸ, ਜੋ ਪੰਜਾਬੀ ਅਤੇ ਹੋਰ ਭਾਰਤੀ ਭਾਸ਼ਾਵਾਂ ਵਿੱਚ ਸੁਪਰ-ਫਾਸਟ ਆਵਾਜ਼ ਪੈਦਾ ਕਰਦਾ ਹੈ..."	▶️ Download WAV
🇮🇳 Marathi (मराठी)	`Charvi` (Young Female, Upbeat)	"नमस्कार! ओरकांता एआय मध्ये आपले सहर्ष स्वागत आहे. सादर करत आहोत ओरकांता व्हॉईस, भारताचे प्रगत रिअल-टाईम ऑडिओ जनरेशन तंत्रज्ञान..."	▶️ Download WAV
🇮🇳 Kannada (ಕನ್ನಡ)	`Nakul` (Young Male, Youthful)	"ನಮಸ್ಕಾರ! ಓರ್ಕಾಂತಾ ಎಐ ಗೆ ಸುಸ್ವಾಗತ. ಭಾರತದ ಅತ್ಯಂತ ವೇಗದ ಮತ್ತು ನೈಸರ್ಗಿಕ ಧ್ವನಿ ತಂತ್ರಜ್ಞಾನವಾದ ಓರ್ಕಾಂತಾ ವಾಯ್ಸ್ ಅನ್ನು ಅನುಭವಿಸಿ..."	▶️ Download WAV

🌟 Key Highlights

12 Languages Supported: English (en), Odia (or), Hindi (hi), Telugu (te), Tamil (ta), Bengali (bn), Marathi (mr), Gujarati (gu), Kannada (kn), Malayalam (ml), Punjabi (pa), and Assamese (as).
Ultra-Low Latency: ~14 ms TTFA on H100 GPU; <90 ms on T4 GPU in Google Colab.
Continuous-Batching CUDA Graphs: Zero CPU overhead during generation loops. One GPU dynamically serves dozens to hundreds of concurrent WebSocket callers.
Studio-Quality 24 kHz Sound: 8-codebook neural acoustic compression powered by Kyutai's Mimi architecture.
Instant Zero-Shot Voice Cloning: Clone any voice with 3–10 seconds of reference speech.
Intelligent Indic Text Normalization (TN): Automatically turns digits (୨୪ ➔ ଚବିଶି), currencies (₹୫୦୦ ➔ પાંચસો રૂપિયા), dates, abbreviations, and percentages into spoken words.
Production-Ready Servers: Includes built-in OpenAI-compatible HTTP /v1/audio/speech endpoint and low-latency WebSocket bidirectional streaming server.

🗣️ Supported Languages & Built-in Voices

Orqanta-Voice ships with 20 built-in named voices across diverse genders, tones, and ages:

Voice Name	Gender	Age / Style	Supported Languages
`Siddhant`	Male	Adult, Clear, Professional	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Vaidehi`	Female	Adult, Calm, Gentle	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Ahana`	Female	Young, Bright, Youthful	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Anirban`	Male	Adult, Steady Narrator	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Malini`	Female	Mature, Steady Narrator	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Nakul`	Male	Young, Energetic, Upbeat	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Charvi`	Female	Young, Upbeat, Dynamic	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Satvik`	Male	Young, Warm, Friendly	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`, `pa`, `as`)
`Jahnavi`	Female	Adult, Warm, Friendly	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`)
`Vasant`	Male	Mature, Authoritative	Multilingual (`en`, `hi`, `or`, `te`, `ta`, `bn`, `mr`, `gu`, `kn`, `ml`)
`Sreekar`	Male	Adult, Natural Conversational	Native Telugu & English/Hindi (`te`, `en`, `hi`)
`Kayal`	Female	Adult, Natural Conversational	Native Tamil & English (`ta`, `en`)
`Abhay`	Male	Adult, Casual Conversational	Native Hindi & English (`hi`, `en`)
`Nivedita`	Female	Adult, Professional News	Native Hindi & English (`hi`, `en`)
`Ishwari`	Female	Adult, Expressive	Native Hindi (`hi`)
`Ritvik`	Male	Adult, Deep Baritone	Native Hindi (`hi`)
`Vrinda`	Female	Young, Gentle	Native Hindi (`hi`)
`Shaunak`	Male	Young, Conversational	Native Hindi & English (`hi`, `en`)
`Akshat`	Male	Adult, Dynamic Storyteller	Native Hindi & English (`hi`, `en`)
`Tanushree`	Female	Adult, Warm Conversational	Native Hindi & English (`hi`, `en`)

🚀 Quickstart

1. One-Click Google Colab (Recommended)

You can run Orqanta-Voice for free on a Google Colab T4 GPU:

!git clone https://github.com/GANDHIKALASI/Orqanta-Voice.git
%cd Orqanta-Voice
!pip uninstall -y torchvision
!pip install -e .
!pip install soundfile

2. Python API Usage

import soundfile as sf
from orqanta_voice import OrqantaVoiceTTS

# Initialize engine on CUDA with 1 generation seat
tts = OrqantaVoiceTTS.from_pretrained(device="cuda", max_seats=1)

# Synthesize speech (Text Normalization enabled by default for numbers and currency)
text = "ଆସନ୍ତା ୨୪ ଘଣ୍ଟା ମଧ୍ୟରେ ପବନର ବେଗ ୧୨୦ କିଲୋମିଟର ହେବାର ସମ୍ଭାବନା ଅଛି।"
waveform = tts.generate(text, lang="or", voice="Vaidehi")

# Save 24 kHz audio
sf.write("output.wav", waveform, 24000)

3. Zero-Shot Voice Cloning

Clone any custom voice in real-time from just 3 to 10 seconds of reference audio:

# 1. Register a custom voice
tts.register_voice(
    name="my_voice",
    ref_audio="my_reference_sample.wav",      # Path to clear audio
    ref_text="The exact words spoken in the audio clip",
    lang="en"
)

# 2. Synthesize using the cloned voice across any language
cloned_wav = tts.generate(
    "Hello! Now I am speaking in your cloned voice with high accuracy.",
    lang="en",
    voice="my_voice"
)
sf.write("cloned_output.wav", cloned_wav, 24000)

4. High-Performance WebSocket Streaming Server

For ultra-low-latency real-time voice bots (audio streaming chunks as they are generated):

# Start the streaming server
python -m orqanta_voice.server --host 0.0.0.0 --port 8000 --seats 8

Stream in Python using the async client:

import asyncio
from orqanta_voice.client import AsyncOrqantaVoiceClient

async def main():
    client = AsyncOrqantaVoiceClient("ws://localhost:8000/v1/stream")
    async for chunk in client.stream_speech("Streaming speech token by token in real time.", lang="en", voice="Siddhant"):
        # Play or forward raw 24 kHz PCM chunk immediately
        pass

asyncio.run(main())

5. OpenAI-Compatible HTTP Server

Integrate Orqanta-Voice directly into any OpenAI TTS client or frontend:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

response = client.audio.speech.create(
    model="orqanta-voice",
    voice="Vaidehi",
    input="नमस्ते! ओरकांता वॉइस में आपका स्वागत है।"
)
response.stream_to_file("hindi_output.mp3")

📊 Technical Architecture & Benchmarks

                       ┌───────────────────────────────────────┐
                       │   Text Input (11 Indic + English)     │
                       └──────────────────┬────────────────────┘
                                          │ Text Normalization (TN)
                                          ▼
                       ┌───────────────────────────────────────┐
                       │  Qwen3 (0.6B) Language Backbone       │
                       │  - Autoregressive Codebook 0 tokens   │
                       │  - Continuous-Batching CUDA Graphs   │
                       └──────────────────┬────────────────────┘
                                          │
                                          ▼
                       ┌───────────────────────────────────────┐
                       │  8-Layer Depth Transformer            │
                       │  - Parallel Codebooks 1-7 Prediction  │
                       │  - Speaker Anchor (192-dim ECAPA)     │
                       └──────────────────┬────────────────────┘
                                          │
                                          ▼
                       ┌───────────────────────────────────────┐
                       │  Mimi Neural Audio Codec (24 kHz)     │
                       │  - Continuous Frame Upsampling        │
                       │  - Hi-Fi 24 kHz Output Waveform       │
                       └───────────────────────────────────────┘

TTFA (Time-to-First-Audio): ~14 ms (H100, B=1)
Concurrent Throughput: 224 simultaneous streams at 323 seconds of audio per real-time second.
P99 Latency: 242 ms under maximum steady-state load with 0 playback stalls.
Memory Footprint: 3.8 GB GPU VRAM (FP16/BF16 weights and KV caches).

🏷️ Model Metadata & Specifications

Property	Details
Model Type	Streaming Autoregressive Neural Text-to-Speech (TTS)
License	Apache-2.0
Supported Languages	English (`en`), Odia (`or`), Hindi (`hi`), Telugu (`te`), Tamil (`ta`), Bengali (`bn`), Marathi (`mr`), Gujarati (`gu`), Kannada (`kn`), Malayalam (`ml`), Punjabi (`pa`), Assamese (`as`)
Acoustic Backbone	Qwen3 (0.6B) + 8-Layer Depth Transformer
Audio Codec	Kyutai Mimi Neural Codec (8 codebooks @ 12.5 Hz framerate)
Sample Rate	24,000 Hz (24 kHz) Hi-Fi Waveform
Speaker Embeddings	ECAPA-TDNN (192-dimensional continuous speaker anchor)
Latency Benchmark	~14 ms TTFA (H100) / <90 ms TTFA (Colab T4 GPU)
Tags	`text-to-speech`, `tts`, `audio`, `speech`, `indic`, `streaming`, `voice-cloning`, `zero-shot`

👨‍💻 Developer & Maintainer

Gandhi Kalasi

AI Product Engineer Aspirant | Creator & Maintainer of Orqanta AI & Orqanta Voice

Passionate about building AI-powered products, exploring large language models (LLMs) and AI agents, and developing practical software solutions that turn complex ideas into useful, real-world applications.

📄 License

Orqanta-Voice is licensed under the Apache License 2.0.

Downloads last month: -