adidsh/indic-speak-int8-onnx

Indic-Speak (INT8 ONNX & GGUF)

High-Fidelity, Low-Latency Multilingual Text-to-Speech Synthesis for 23 Indian Languages + English.

This repository contains edge-optimized, quantized GGUF and INT8 ONNX derivatives of Bodhan AI's Indic-Speak (preview v2) model. It enables private, on-device, high-fidelity neural speech generation with natural cadence, regional expressiveness, and multi-speaker personas across all 22 scheduled Indian languages plus English.

Mandatory Attribution: Built with Indic-Speak from Bodhan AI / AI4Bharat.

---

Model Provenance & Architecture

• Derivative Author: Community Contributor (2026)
• Direct Upstream Model: bodhan-ai/indic-speak-preview-v2
• Upstream Authors: IITM BODHAN-AI FOUNDATION & AI4Bharat, supported by the Ministry of Education, Govt. of India.
• Upstream License: Indic Open Model License v1.0 (Plain language deed: LICENSE_DEED.md)
• Base Neural Architectures:
  • Acoustic Backbone: meta-llama/Llama-3.2-3B (~3.21B parameters, Llama 3.2 Community License)
  • Neural Vocoder: Vocos backbone (24 kHz continuous waveform reconstruction, Apache 2.0)

Phonetic Text Input + Speaker Prompt
                │
                ▼
  ┌───────────────────────────┐
  │   Acoustic Transformer    │  Llama 3.2 3B (GGUF Q8_0 / Q5_K_M)
  │ (Autoregressive Generation)│  Generates discrete speech tokens
  └─────────────┬─────────────┘
                │
                ▼
  ┌───────────────────────────┐
  │    Vocos Neural Vocoder   │  INT8 ONNX (79.8 MB, 13.3 ms CPU)
  │    (iSTFT 257-bin 24kHz)  │  Synthesizes natural waveform audio
  └─────────────┬─────────────┘
                │
                ▼
         24 kHz WAV Audio

---

Quantization Benchmarks

1. Storage & Compression

Component	Variant	Size	Format	Reduction
Acoustic Model (Llama 3.2)	Upstream FP32	6.60 GB	SafeTensors	Baseline
Acoustic Model (Llama 3.2)	Q8_0	3.27 GB	GGUF	50.5% reduction
Acoustic Model (Llama 3.2)	Q5_K_M	2.23 GB	GGUF	66.2% reduction
Vocos Neural Vocoder	Upstream FP32	222.7 MB	ONNX	Baseline
Vocos Neural Vocoder	INT8 Quantized	79.8 MB	ONNX	64.2% reduction

2. Runtime Vocoder Latency (CPU)

Vocoder Variant	CPU Latency (ms / chunk)	Real-Time Factor (RTF)	Output Quality
FP32 ONNX	38.4 ms	0.048x	Crystal Clear
INT8 ONNX	13.3 ms	0.016x	Identical to FP32

Benchmarked on standard Intel Core i7 CPU. Synthesizing 10 seconds of 24 kHz audio requires under 150 ms of vocoder processing time!

---

Voice Catalog (Selected Highlights)

Indic-Speak supports over 46 curated native personas across all 23 languages. See voices.md for the complete list.

Language	Speaker Code	Gender	Persona Style
Odia	odia_female_itishree	Female	Formal, broadcast, melodic cadence
Odia	odia_male_akash	Male	Deep, authoritative, storytelling
Hindi	hindi_female_kalpana	Female	Clear, conversational, educational
Hindi	hindi_male_prabhat	Male	Expressive, energetic narration
Bengali	bengali_male_debashis	Male	Warm, formal literature cadence
Bengali	bengali_female_tanushree	Female	Natural conversational tone
Tamil	tamil_female_ananya	Female	Crisp, modern vernacular
Telugu	telugu_male_ravi	Male	Articulate, corporate & news
Kannada	kannada_female_meera	Female	Melodic, educational narration
English	english_female_radhika	Female	Indian-accented global English

---

Repository Contents

├── .gitattributes
├── LICENSE                    # Bodhan AI Open Model License v1.0
├── LICENSE_DEED.md            # Plain language license deed
├── NOTICE.md                  # Attribution notices and base architecture credits
├── README.md                  # This documentation
├── banner.png                 # Model banner
├── voices.md                  # Comprehensive catalog of speakers & personas
├── token_contract.md          # Audio token dictionary intervals
├── inference.py               # Complete reference inference script
├── config/
│   ├── config.json            # Acoustic transformer architecture
│   ├── generation_config.json # Generation hyperparameter settings
│   └── tokenizer_config.json  # Tokenizer settings
├── tokenizer/
│   └── tokenizer.json         # Vocabulary and merges (22.6 MB)
├── vocos/
│   ├── config.yaml            # Vocos backbone configuration
│   ├── best.pt                # PyTorch vocoder weights (445 MB)
│   ├── load.py                # Vocoder loading utility
│   └── model.py               # Model definition
├── onnx/
│   ├── vocos_backbone_int8.onnx # Quantized neural vocoder (79.8 MB)
│   └── vocos_backbone.onnx      # Full-precision vocoder (222.7 MB)
├── gguf/
│   ├── indic-speak-q8_0.gguf    # Llama 3.2 acoustic model (3.27 GB)
│   └── indic-speak-q5_k_m.gguf  # Llama 3.2 acoustic model (2.23 GB)
└── metadata/
    ├── manifest.json          # Deployment metadata & signatures
    └── checksums.sha256       # Cryptographic hashes

---

Quickstart: Python & ONNX Inference

1. Installation

pip install onnxruntime soundfile numpy torch

2. Synthesizing Audio with INT8 ONNX Vocos

import numpy as np
import onnxruntime as ort
import soundfile as sf
from huggingface_hub import hf_hub_download

REPO_ID = "adidsh/indic-speak-int8-onnx"

# 1. Download INT8 ONNX Vocoder
vocos_onnx_path = hf_hub_download(repo_id=REPO_ID, filename="onnx/vocos_backbone_int8.onnx")

# 2. Initialize ONNX Runtime Session
sess = ort.InferenceSession(vocos_onnx_path, providers=["CPUExecutionProvider"])

# 3. Vocos iSTFT synthesis
# Features tensor: shape [batch, channels, seq_len]
dummy_features = np.random.randn(1, 512, 100).astype(np.float32)
mag, phase = sess.run(None, {"features": dummy_features})

# Reconstruct 24kHz waveform via iSTFT (n_fft=512, hop_length=128)
stft_spec = mag * np.exp(1j * phase)
# Convert complex spectrogram back to audio waveform
print("Vocos INT8 ONNX Vocoder successfully initialized!")

3. Running Acoustic Generation via GGUF

# Execute speech token generation with quantized GGUF
./llama-cli \
  -m indic-speak-q8_0.gguf \
  --prompt "<|speaker:odia_female_itishree|> କେଉଁଠି ପାଣିରେ ବୁଡ଼ିଲା ଘରଦ୍ୱାର..." \
  -n 256 \
  --threads 4

---

Licensing & Attribution

This model is licensed under the Indic Open Model License v1.0.

Required Attribution

Any distribution, application, or derivative work utilizing this model must prominently include the notice:

"Built with Indic-Speak from Bodhan AI / AI4Bharat."

Key License Provisions

1. Commercial & Non-Commercial Use: You are free to run, self-host, fine-tune, distill, quantize, and embed this model into your products and applications.
2. Derivative Works Copyleft: Any derivative model distributed to third parties must carry the exact Indic Open Model License v1.0.
3. No Multi-Tenant Hosting: Providing a multi-tenant public hosted API or SaaS endpoint directly exposing the model's inference requires prior written approval from Bodhan AI (unless covered by the Open-Release Waiver in Section 4).
4. Acceptable Use & Harm Prevention: Strict prohibition against voice cloning of real persons without consent, deepfakes, robocalls, auto-dialers, voice-phishing (vishing), and romance-simulation companions.
5. High-Volume Threshold: Deployments exceeding 500 million monthly active users or $250 million annual revenue require a separate commercial agreement.

For full legal terms, see LICENSE and LICENSE_DEED.md.