Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/call_sample_1.wav filter=lfs diff=lfs merge=lfs -text
+examples/call_sample_2.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,20 @@
 ---
-title: Nepaliconformer
-emoji: 🏢
-colorFrom: green
-colorTo: indigo
+title: NepaliConformer
+emoji: 🏔️
+colorFrom: red
+colorTo: gray
 sdk: gradio
-sdk_version: 6.24.0
-python_version: '3.13'
+sdk_version: "5.44.1"
 app_file: app.py
 pinned: false
+license: cc-by-nc-4.0
+short_description: Nepali ASR built for real telephone calls, not read speech
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# NepaliConformer
+
+Transcribe Nepali speech with a model trained for **real telephone audio**. On the NepTel
+real-call benchmark it scores 33.8% WER where Whisper-large-v3 zero-shot scores ~99%.
+
+Full release — models, benchmark, honest results including every measured limitation:
+**https://github.com/Ampixa/nepaliconformer**

app.py

@@ -0,0 +1,59 @@
+"""NepaliConformer demo — transcribe Nepali speech, tuned for real telephone audio."""
+import os
+
+import gradio as gr
+import torch
+
+MODEL = None
+
+
+def get_model():
+    global MODEL
+    if MODEL is None:
+        from huggingface_hub import hf_hub_download
+        from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
+        path = hf_hub_download("ampixa/nepali-conformer-offline",
+                               "nepali_conformer_offline.nemo")
+        torch.set_num_threads(2)
+        MODEL = EncDecHybridRNNTCTCBPEModel.restore_from(path, map_location="cpu")
+        MODEL.eval()
+    return MODEL
+
+
+def transcribe(audio_path):
+    if not audio_path:
+        return "(no audio)"
+    import soundfile as sf
+    info = sf.info(audio_path)
+    if info.duration > 60:
+        return "Please keep clips under 60 seconds for this CPU demo."
+    model = get_model()
+    out = model.transcribe([audio_path], batch_size=1, verbose=False)[0]
+    text = out.text if hasattr(out, "text") else str(out)
+    return " ".join(t for t in text.split() if t != "<breath>") or "(no speech detected)"
+
+
+examples = [[f"examples/{f}"] for f in sorted(os.listdir("examples"))] \
+    if os.path.isdir("examples") else None
+
+demo = gr.Interface(
+    fn=transcribe,
+    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath",
+                    label="Nepali speech (mic or file, ≤60 s)"),
+    outputs=gr.Textbox(label="Transcript (Devanagari)", rtl=False),
+    title="NepaliConformer — Nepali ASR for real telephone calls",
+    description=(
+        "121M Conformer trained on ~1,655 h of conversational Nepali. "
+        "**33.8% WER on real call-center audio** (NepTel benchmark) where Whisper-large-v3 "
+        "zero-shot scores ~99%. Honest limitations and the full benchmark: "
+        "[github.com/Ampixa/nepaliconformer](https://github.com/Ampixa/nepaliconformer). "
+        "CPU demo — a 30 s clip takes roughly 10-20 s. Example clips are real call-center "
+        "audio (CC-BY-4.0, © InfoBayAI)."
+    ),
+    examples=examples,
+    cache_examples=False,
+    flagging_mode="never",
+)
+
+if __name__ == "__main__":
+    demo.launch()

examples/call_sample_1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e300320ba159319392cdae2b0f6c0573a1a4ee4498dc41486f8ef49a2617e84f
+size 798764

examples/call_sample_2.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29cb08bbb62b5dcfd3d5aa454fc535652a338d7aaedd998bab35308f8f24d6d2
+size 796844

requirements.txt

@@ -0,0 +1,4 @@
+nemo_toolkit[asr]==2.3.1
+torch>=2.3,<2.6
+soundfile
+huggingface_hub