Initial commit for asr-demo

.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
+*.arpa filter=lfs diff=lfs merge=lfs -text
+*.binary filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,62 @@
+import gradio as gr
+from transformers import AutoProcessor, AutoModelForCTC
+from pyctcdecode import build_ctcdecoder
+import librosa
+import torch
+
+# Load models and LM files from local files (uploaded to Space)
+processor = AutoProcessor.from_pretrained("ai4bharat/indicwav2vec-hindi")
+model = AutoModelForCTC.from_pretrained("ai4bharat/indicwav2vec-hindi")
+
+vocab = processor.tokenizer.get_vocab()
+sorted_vocab = sorted(vocab.items(), key=lambda kv: kv[1])
+tokens = [token for token, _ in sorted_vocab]
+
+arpa_pruned = "./hindi_small_4gram_pruned_clean.arpa"
+binary_pruned = "./hindi_small_4gram_pruned.binary"
+
+def load_unigrams(arpa_path):
+    unigrams = {}
+    with open(arpa_path, encoding="utf-8") as f:
+        in_unigrams = False
+        for line in f:
+            line = line.strip()
+            if not in_unigrams:
+                if line == "\\1-grams:":
+                    in_unigrams = True
+                continue
+            if line.startswith("\\"):
+                break
+            parts = line.split(maxsplit=2)
+            if len(parts) >= 2:
+                score = float(parts[0])
+                tok   = parts[1]
+                unigrams[tok] = score
+    return unigrams
+
+unigram_scores = load_unigrams(arpa_pruned)
+decoder = build_ctcdecoder(tokens, binary_pruned, unigrams=unigram_scores)
+
+def transcribe(audio):
+    if audio is None:
+        return "No audio provided."
+    sr, audio_np = audio
+    if sr != 16000:
+        audio_np = librosa.resample(audio_np, orig_sr=sr, target_sr=16000)
+        sr = 16000
+    inputs = processor(audio_np, sampling_rate=sr, return_tensors="pt")
+    with torch.no_grad():
+        logits = model(**inputs).logits.cpu().numpy()[0]
+    text = decoder.decode(logits)
+    return text
+
+iface = gr.Interface(
+    fn=transcribe,
+    inputs=gr.Audio(sources=["microphone", "upload"], type="numpy", label="Record or Upload Audio"),
+    outputs="text",
+    title="Indic ASR Demo (Hindi)",
+    description="Record or upload Hindi audio. Get instant transcription using ASR + custom Language Model."
+)
+
+if __name__ == "__main__":
+    iface.launch()

hindi_small_4gram_pruned.binary

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd0483a6dc50e3baf988d5858f966578c9532e402a7128f29590e02e1180d800
+size 282972866

hindi_small_4gram_pruned_clean.arpa

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96d44938524ca3ae7fb7a1470fc0dc3e11c2f6ecf0022ed2b6982f01b5d43dd1
+size 535622088

requirements.txt

@@ -0,0 +1,5 @@
+transformers
+pyctcdecode
+torch
+librosa
+gradio