svara-TTS v1 — Open Multilingual TTS for India’s Voices
svara-TTS is a developer-first multilingual TTS model for 19 languages (18 Indic + Indian English).
Built on an Orpheus-style discrete audio token approach, it targets clarity, expressiveness, and low-latency on commodity GPUs/CPUs.
It supports light-weight emotion/style control (e.g., <happy>, <sad>, <anger>, <fear>) and simple speaker identities (Language (Gender)), with zero-shot adaptation paths.
At a Glance
- Languages (19): Hindi, Bengali, Marathi, Telugu, Kannada, Bhojpuri, Magahi, Chhattisgarhi, Maithili, Assamese, Bodo, Dogri, Gujarati, Malayalam, Punjabi, Tamil, Nepali, Sanskrit, Indian English.
- Expressivity: End-of-utterance style tags; natural prosody; code-switch aware.
- Latency & Deployment: Works well with GGUF exports; suitable for edge/CPU scenarios.
- Adaptability: LoRA-friendly for quick speaker/domain specialization.
Try it live on the
Demo Space, or on
Colab
Deployment scripts and
inference repo will be available soon. Watch our
Github for updates
Prompting (Orpheus-style)
- Place style/emotion tags at the end of the sentence:
आज... सच में अच्छी खबर है — शाम को मिलते हैं! <happy>
- Use punctuation to hint prosody (ellipses, commas, exclamation).
- For technical or dense text, end with
<clear> to prioritize intelligibility.
Speaker IDs follow a simple convention: Language (Gender) (e.g., Marathi (Male)).
Training Data Summary
Trained on 2000+ hours of open, high-quality speech from SYSPIN, RASA, IndicTTS, and SPICOR, covering ~50 speakers (balanced male/female) across 19 languages.
Data was curated to encourage natural prosody, broad coverage, and stable multilingual transfer. See Acknowledgments for provenance.
Intended Uses
- Multilingual assistants, IVR, learning apps, reading aids, accessibility tools
- Content localization (education, public-information, civic services)
- Research on Indic prosody, emotion control, cross-lingual transfer
Out-of-Scope / Not Intended
- Impersonation of private individuals or public figures without consent
- Deceptive content (fraud, harassment, misinformation)
- Safety-critical deployments without human oversight
Limitations
- Proper nouns & rare entities: may require spelling hints or
<clear>.
- Very long sentences: chunk or add punctuation for natural prosody.
- Emotion strength: varies by language due to data density.
- Code-mixing: common patterns work; it’s not a deterministic rules engine.
Many of these improve with targeted LoRA finetuning and better preprocessing.
Responsible Use
By using this model, you agree to follow applicable laws and ethical guidelines.
Avoid impersonation, harassment, targeted deception, or other harmful uses.
Where appropriate, disclose synthetic speech to end users.
Sources & Links
🙏 Acknowledgments
This work was developed by
Kenpath Technologies for the open-source community. We also thank RunPod for the startup credits that supported our GPU compute.
- Canopy Labs — Orpheus: foundational ideas & open release
Release: https://canopylabs.ai/releases/orpheus_can_speak_any_language
- SPIRE Lab, IISc Bangalore — SYSPIN (multilingual studio) and SPICOR (Indian English)
- AI4Bharat — RASA expressive speech
- IIT Madras — IndicTTS
- Unsloth — helpful notes & tooling
- RunPod — startup GPU credits that accelerated experiments
License
Apache-2.0
Versioning & Changelog
- v1.0.0: Initial public release (19 languages)
