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
Download code/export/extract_tower_rest.py from junafinity/VeriSafe-Tanglish-1-STT: direct link, hf CLI and curl.
- Browser
- Download file 1.32 kB
-
https://huggingface.co/junafinity/VeriSafe-Tanglish-1-STT/resolve/main/code/export/extract_tower_rest.py
- Command line
-
hf download hf://junafinity/VeriSafe-Tanglish-1-STT/code/export/extract_tower_rest.py
-
curl -L -o extract_tower_rest.py https://huggingface.co/junafinity/VeriSafe-Tanglish-1-STT/resolve/main/code/export/extract_tower_rest.py
1.32 kB
| """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}") | |