nullHawk/snac-24khz-hindi-ft

SNAC 24 kHz — Hindi

A SNAC 24 kHz multi-scale neural audio codec wh ose decoder has been fine-tuned on Hindi / Hinglish speech. It serves as the vocoder (SNAC-token → waveform detokenizer) for a Orpheus styled TTS stack.

Only the decoder was fine-tuned — the encoder and quantizer are frozen from the base hubertsiuzdak/snac_24khz. The code space is therefore identical to the base model, so this checkpoint is a drop-in decoder replacement: any SNAC code24khz` decode here with no retraining and no re-tokenizing.

Model details

Base	`hubertsiuzdak/snac
Output	24 kHz, mono
Codebooks	3 hierarchical levels, 4096 entries each
Frame layout	7 tokens / ine)
Parameters	~19.8 M
Fine-tuned	decoder only

Installation

pip install snac torch sound

Usage

1. Round-trip: audio → c

import torch, soundfile as s
from snac import SNAC

device = "cuda" if torch.cud
model = SNAC.from_pretrained("nullHawk/snac-24khz-hindi-hp").eval().to(device)

wav, sr = sf.read("input.wavt be 24 kHz (resample first if
not)
audio = torch.from_numpy(wavce)   # shape [1, 1, T]

with torch.inference_mode():
    codes = model.encode(aud LongTensors (coarse -> fine)
    recon = model.decode(codes)               # [1, 1, T'] waveform

sf.write("recon.wav", recon[

2. Decode from a flat SN

TTS LLMs (Orpheus SNAC tokens, 7 per frame, interleaved across the 3 lev:

import torch

def decode_snac_tokens(ids,
    """ids   : flat list of
       offset: subtract your LLM's SNAC base id (use 0 if ids are already in [0, 4096)).
"""
    ids = [(t - offset) % 40
    frames = len(ids) // 7
    l1, l2, l3 = [], [], []
    for i in range(frames):
        s = ids[i * 7:(i + 1) * 7]
        l1.append(s[0])
        l2 += [s[1], s[4]]
        l3 += [s[2], s[3], s
    dev = next(model.parameters()).device
    codes = [torch.tensor(x,)[None] for x in (l1, l2, l3)]
    with torch.inference_mode():
        z_q = model.quantize
        audio = model.decoder(z_q)[0, 0].cpu().numpy()
    return audio[2048:]      ms decoder warm-up

# offset = the SNAC base tokg. 128266 for maya1 / Orpheus).
wav = decode_snac_tokens(token_ids, model, offset=128266)
import soundfile as sf
sf.write("tts.wav", wav, 24000)

Notes

• Only the decoder changed, tsiuzdak/snac_24khz` by swappin g the repo id — the encoded code
• Output is 24 kHz mono. Trim the first ~2048 samples of a freshly decoded clip to remov e the decoder warm-up transient.

Acknowledgements

Built on SNAC by Hubert Siuzdak.