Automatic Speech Recognition
PyTorch
Safetensors
GGUF
MLX
Tamil
English
speech
asr
streaming
rnnt
ctc
tamil
tanglish
code-switching
indian-english
verisafe
Instructions to use junafinity/VeriSafe-Tanglish-1-STT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use junafinity/VeriSafe-Tanglish-1-STT with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download junafinity/VeriSafe-Tanglish-1-STT --local-dir VeriSafe-Tanglish-1-STT
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
File size: 1,324 Bytes
ec01e2d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | """Write the only EmbeddingGemma 2 tensors the stage 6 model needs that best.pt does not hold.
best.pt saves trainable parameters only (train.py weights_state). 482 audio-tower tensors stay behind in the base model: 480 clip
bounds (input_min/max, output_min/max of each Gemma4ClippableLinear, finite, used in forward) and output_proj (unused by the STT
path, kept so a strict load passes). This writes them as float32, the dtype the project's from_pretrained(dtype=float32) uses.
python extract_tower_rest.py EG2_MODEL.safetensors BEST.pt OUT.safetensors
"""
import sys
import torch
from safetensors import safe_open
from safetensors.torch import save_file
src, ckpt, out = sys.argv[1:4]
held = {k[len("tower."):] for k in torch.load(ckpt, map_location="cpu", weights_only=False, mmap=True)["model"] if k.startswith("tower.")}
rest = {}
with safe_open(src, "pt") as f:
for k in f.keys():
if k.startswith("audio_tower.") and k[len("audio_tower."):] not in held:
rest["tower." + k[len("audio_tower."):]] = f.get_tensor(k).float().clone()
save_file(rest, out, metadata={"source": "google/embeddinggemma-2 snapshot f9972ac22e696a105c482202646979e33e4b5a1e, audio_tower tensors absent from stage 6 best.pt"})
print(f"{len(rest)} tensors, {sum(t.numel() for t in rest.values())} values -> {out}")
|