kenpath/svara-tts-v1-openvino-int4

Svara TTS v1 -- OpenVINO INT4

High-Quality Multilingual Indic Text-to-Speech on Intel Hardware

---

Overview

Svara TTS v1 is a 3.3B-parameter autoregressive text-to-speech model built on the LLaMA architecture, fine-tuned to generate natural-sounding speech across 19 Indic languages and Indian English. It uses the SNAC neural audio codec with 7-band interleaved codes to produce high-fidelity 24 kHz audio.

This is the OpenVINO INT4 quantized version, optimized for deployment on Intel hardware including CPUs, integrated GPUs, and NPUs (e.g., Intel Core Ultra / Panther Lake).

Architecture	LlamaForCausalLM (3.3B params)
Quantization	INT4 (OpenVINO NNCF)
Model Size	1.94 GB (down from ~6.6 GB in BF16)
Audio Codec	SNAC (7 bands, 4096 audio vocab)
Sample Rate	24,000 Hz
Max Context	131,072 tokens
Optimum Version	2.1.0

---

Supported Languages

Language	Code	Speakers	Language	Code	Speakers
Hindi	hi	Male, Female	Malayalam	ml	Male, Female
Bengali	bn	Male, Female	Punjabi	pa	Male, Female
Tamil	ta	Male, Female	Assamese	as	Male, Female
Telugu	te	Male, Female	Nepali	ne	Male, Female
Marathi	mr	Male, Female	Bhojpuri	bh	Male, Female
Gujarati	gu	Male, Female	Dogri	dg	Male, Female
Kannada	kn	Male, Female	Sanskrit	sa	Male, Female
Indian English	en	Male, Female	Maithili	mt	Male, Female
Chhattisgarhi	cc	Male, Female	Bodo	bo	Male, Female
Magahi	mg	Male, Female

---

Quick Start

Installation

pip install optimum[openvino] transformers torch snac soundfile

Basic Usage

from optimum.intel import OVModelForCausalLM
from transformers import AutoTokenizer

model_id = "kenpath/svara-tts-v1-openvino-int4"

# Load model -- choose your Intel device
model = OVModelForCausalLM.from_pretrained(model_id, device="CPU")  # or "GPU", "NPU"
tokenizer = AutoTokenizer.from_pretrained(model_id)

Full Text-to-Speech Pipeline

import torch
import numpy as np
import soundfile as sf
from snac import SNAC

# ── Special tokens ──
BOS_TOKEN      = 128000
END_OF_TURN    = 128009
START_OF_SPEECH = 128257
END_OF_SPEECH  = 128258
START_OF_HUMAN = 128259
END_OF_HUMAN   = 128260
START_OF_AI    = 128261
AUDIO_TOKEN    = 156939
AUDIO_TOKENS_START = 128266
BAND_OFFSETS   = [128266, 132362, 136458, 140554, 144650, 148746, 152842]

# ── Build input ──
speaker = "hi_female"   # format: {lang_code}_{male|female}
text = "नमस्ते, आज का मौसम बहुत अच्छा है।"

prompt_text = f"{speaker}: {text}"
text_ids = tokenizer.encode(prompt_text, add_special_tokens=False)
input_ids = torch.tensor([[
    BOS_TOKEN, START_OF_HUMAN, AUDIO_TOKEN,
    *text_ids,
    END_OF_HUMAN, END_OF_TURN, START_OF_AI, START_OF_SPEECH
]])

# ── Generate audio tokens ──
output = model.generate(
    input_ids=input_ids,
    max_new_tokens=4096,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    repetition_penalty=1.1,
)

# ── Decode with SNAC ──
generated = output[0][input_ids.shape[1]:]
audio_tokens = [t.item() for t in generated if t.item() >= AUDIO_TOKENS_START]
audio_tokens = audio_tokens[:len(audio_tokens) // 7 * 7]  # align to 7-band frames

codes_0, codes_1, codes_2 = [], [], []
for i in range(0, len(audio_tokens), 7):
    frame = audio_tokens[i:i+7]
    codes_0.append(frame[0] - BAND_OFFSETS[0])
    codes_1.extend([frame[1] - BAND_OFFSETS[1], frame[4] - BAND_OFFSETS[4]])
    codes_2.extend([
        frame[2] - BAND_OFFSETS[2], frame[3] - BAND_OFFSETS[3],
        frame[5] - BAND_OFFSETS[5], frame[6] - BAND_OFFSETS[6],
    ])

snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval()
with torch.no_grad():
    audio = snac.decode(torch.tensor([codes_0, codes_1, codes_2]))
audio_np = audio.squeeze().cpu().numpy()

sf.write("output.wav", audio_np, 24000)
print(f"Saved output.wav ({len(audio_np)/24000:.1f}s)")

---

Speaker Format

Speakers follow the pattern {language_code}_{gender}:

hi_female    # Hindi female
ta_male      # Tamil male
en_female    # Indian English female
bn_male      # Bengali male

---

Intel Hardware Targets

Device	Flag	Best For
CPU	device="CPU"	Universal compatibility
Integrated GPU	device="GPU"	Balanced performance
NPU	device="NPU"	Intel Core Ultra / Panther Lake -- lowest power

INT4 quantization provides ~3.4x compression vs the original BF16 weights while maintaining speech quality across all supported languages.

---

Model Architecture

LlamaForCausalLM (Modified for TTS)
├── Embedding: 156,940 tokens (text + 28,674 audio tokens)
├── 28 x Transformer Blocks
│   ├── Grouped-Query Attention (24 heads, 8 KV heads, dim 128)
│   ├── RMSNorm (eps=1e-5)
│   └── SwiGLU MLP (3072 → 8192 → 3072)
├── RoPE (LLaMA3-style, θ=500K, 128K context)
└── Tied LM Head → 156,940 logits
         ↓
    SNAC Decoder (24 kHz, 7-band, 4096 codebook)
         ↓
      Waveform

---

Files

File	Size	Description
openvino_model.bin	1.94 GB	INT4 quantized weights
openvino_model.xml	2.0 MB	OpenVINO IR graph definition
openvino_tokenizer.bin	11 MB	Compiled tokenizer
openvino_tokenizer.xml	29 KB	Tokenizer graph
openvino_detokenizer.bin	2.6 MB	Compiled detokenizer
openvino_detokenizer.xml	14 KB	Detokenizer graph
tokenizer.json	22 MB	HF tokenizer (156,940 vocab)
config.json	893 B	Model configuration

---

Citation

@misc{svara-tts-v1-openvino-int4,
  title   = {Svara TTS v1 -- OpenVINO INT4},
  author  = {kenpath},
  year    = {2026},
  url     = {https://huggingface.co/kenpath/svara-tts-v1-openvino-int4},
  note    = {INT4 quantized OpenVINO conversion of kenpath/svara-tts-v1}
}

---

Base Model | OpenVINO Docs | SNAC Codec