Veena TTS

Veena – Text-to-Speech Model

What is Veena?

Veena, developed by Maya Research, is a state-of-the-art text-to-speech (TTS) model built on a 3 billion-parameter Llama-based autoregressive transformer. It delivers natural, expressive speech in Hindi and English—handling mixed-language inputs seamlessly. Leveraging the SNAC neural codec at 24 kHz, Veena generates studio-quality audio with four distinct speaker personas (Kavya, Agastya, Maitri, Vinaya). Optimized for ultra-low latency (sub-80 ms on high-end GPUs) and production-ready deployment via 4-bit quantization, Veena is engineered for real-time applications in accessibility, customer service, content creation, and voice-enabled devices.

Key Features

• •High-Fidelity Audio: 24 kHz sampling rate with SNAC neural codec for crystal-clear voice output
• •Multilingual & Code-Switching: Fluent in Hindi and English; natural transitions in mixed-language text
• •Four Unique Voices:
  • •Kavya (warm, friendly)
  • •Agastya (deep, authoritative)
  • •Maitri (clear, neutral)
  • •Vinaya (bright, youthful)
• •Low Latency: Sub-80 ms response time on top-tier GPUs—ideal for live interactions
• •Efficient Quantization: 4-bit precision reduces memory footprint without compromising quality
• •Transformer-Based: 3 billion parameters capture complex intonation, stress, and pacing patterns

Best Use Cases

• •Accessibility Tools: Screen readers, assistive communication devices
• •Customer Service: Interactive voice response (IVR), chatbots, automated agents
• •Content Creation: Podcasts, e-learning narrations, audiobooks
• •Voice-Enabled Devices: Smart speakers, wearables, IoT interfaces
• •Multilingual Platforms: Apps requiring seamless Hindi-English dialogue

Prompt Tips and Output Quality

• •Input Text: For clarity, use simple, declarative sentences; combine complex phrases for emotional nuance.
• •Speaker Selection (speaker):
  • •Default “kavya” for a warm, conversational tone
  • •Switch to “agastya” for a more commanding presence
• •Advanced Controls:
  • •temperature (0–2): 0.2 for monotone, 0.7 for lively expressiveness
  • •top_p (0–1): 0.5 for focused delivery, 0.95 for varied intonation
  • •repetition_penalty (1–2): 1.05 default; increase to 1.2 to minimize repeats
• •Audio Quality: Adjust sampling rate and codec settings for bandwidth or storage constraints without losing clarity

# Veena TTS: Effective Usage Guide

Veena is a high-fidelity, bilingual Hindi–English text-to-speech model with four distinct speaker personas. This guide walks you through getting the best audio quality and response times by tuning core parameters and selecting voice styles for different applications.

## 1. Getting Started
1. **Input Text**  
   – Keep sentences clear and declarative for consistent pacing.  
   – Inject emotive or domain-specific terms to convey nuance.  
2. **Speaker Persona**  
   – **kavya**: Warm, friendly – ideal for podcasts, e-learning  
   – **agastya**: Deep, authoritative – best for IVR, announcements  
   – **maitri**: Neutral, clear – suited for screen readers, accessibility  
   – **vinaya**: Bright, youthful – perfect for marketing, tutorials  

## 2. Core Parameters
| Parameter           | Range      | Default | Effect                                            |
|---------------------|------------|---------|---------------------------------------------------|
| **temperature**     | 0.0 – 2.0  | 0.4     | Controls expressiveness; lower=flat, higher=lively |
| **top_p**           | 0.0 – 1.0  | 0.9     | Nucleus sampling; lower=focused, higher=diverse    |
| **repetition_penalty** | 1.0 – 2.0| 1.05    | Penalizes repeats; higher=fewer looped phrases      |

## 3. Use-Case Presets
1. **Interactive Voice Response (IVR)**  
   – speaker: **agastya**  
   – temperature: **0.2** (steady, formal)  
   – top_p: **0.5** (focused clarity)  
   – repetition_penalty: **1.1**  

2. **Audio Narration / Audiobooks**  
   – speaker: **kavya** or **maitri**  
   – temperature: **0.6** (natural variation)  
   – top_p: **0.95** (rich intonation)  
   – repetition_penalty: **1.05**  

3. **Accessibility & Screen Readers**  
   – speaker: **maitri**  
   – temperature: **0.2** (consistent pacing)  
   – top_p: **0.6** (reliable phrasing)  
   – repetition_penalty: **1.2**  

4. **Marketing & E-Learning Clips**  
   – speaker: **vinaya**  
   – temperature: **0.7** (energetic)  
   – top_p: **0.9** (varied tone)  
   – repetition_penalty: **1.05**  

## 4. Tips for Optimal Quality
- **Code-Switching**: Embed Hindi and English naturally; Veena handles mixed scripts seamlessly.  
- **Sentence Length**: Break very long sentences into shorter ones for stable prosody.  
- **Pauses & Punctuation**: Use commas, ellipses, and line breaks to guide breath and rhythm.  
- **Latency**: On high-end GPUs, expect <80 ms per synthesis; quantized mode reduces memory footprint.

## 5. Troubleshooting
- **Choppy Output**: Lower temperature or top_p.  
- **Monotone Speech**: Increase temperature to 0.6–0.8.  
- **Repetition Artifacts**: Bump repetition_penalty up to 1.2.

By tuning these parameters and choosing the right persona, you can deploy Veena in real-time chatbots, immersive audiobooks, or assistive devices—delivering natural and engaging speech every time.

Can Veena handle Hindi-English code-switching?

Yes. Veena’s transformer backbone is trained on mixed-language corpora for seamless transitions.

What latency should I expect in production?

On high-end GPUs, Veena delivers sub-80 ms end-to-end latency—perfect for real-time use.

How do I pick the best speaker voice?

Choose based on your brand or application tone: Kavya for warmth, Agastya for depth, Maitri for neutrality, Vinaya for energy.

Is a quantized version available?

Absolutely. Veena supports 4-bit quantization for reduced memory usage and faster inference.

What sample rate does Veena output?

Audio is synthesized at 24 kHz using the SNAC neural codec for smooth, high-quality playback.