Update app.py

app.py

@@ -1,65 +1,157 @@
+Of course. I'll update the code to perform automatic language identification based on the transcription's characters and common words before providing the final, high-quality transcription.
+
+This new version will:
+
+1.  **Remove the language dropdown**, as the language will be detected automatically.
+2.  Perform a quick, initial transcription using Hindi as a "pivot" language.
+3.  Analyze the resulting text against a **custom dictionary** of unique characters and common words for all 22 supported languages.
+4.  Once the language is identified, it will perform the final, more accurate transcription using the detected language code.
+
+-----
+
+### **Updated Code with Automatic Language Identification**
+
+Here is the complete, updated code. You can replace your existing script with this one.
+
+```python
 from __future__ import annotations
 import torch
 import torchaudio
 import gradio as gr
 import spaces
 from transformers import AutoModel
+import re
 
-DESCRIPTION = "IndicConformer-600M Multilingual ASR (CTC + RNNT)"
+DESCRIPTION = "IndicConformer-600M Multilingual ASR (CTC + RNNT) with Auto Language ID"
 
-LANGUAGE_NAME_TO_CODE = {
-    "Assamese": "as", "Bengali": "bn", "Bodo": "br", "Dogri": "doi",
-    "Gujarati": "gu", "Hindi": "hi", "Kannada": "kn", "Kashmiri": "ks",
-    "Konkani": "kok", "Maithili": "mai", "Malayalam": "ml", "Manipuri": "mni",
-    "Marathi": "mr", "Nepali": "ne", "Odia": "or", "Punjabi": "pa",
-    "Sanskrit": "sa", "Santali": "sat", "Sindhi": "sd", "Tamil": "ta",
-    "Telugu": "te", "Urdu": "ur"
+# --- Language Identification Data ---
+# A dictionary containing unique character sets and common words for each language.
+# This data is used by our custom language identification logic.
+LANGUAGE_DATA = {
+    "as": {"chars": set("অআইঈউঊঋএঐওঔকখগঘঙচছজঝঞটঠডঢণতথদধনপফবভমযৰলৱশষসহৎংঃঽািীুূৃেৈোৌ্"), "words": set(["আৰু", "হয়", "এটা", "কৰি", "ওপৰত"])},
+    "bn": {"chars": set("অআইঈউঊঋএঐওঔকখগঘঙচছজঝঞটঠডঢণতথদধনপফবভমযরলশষসহৎংঃঽািীুূৃেৈোৌ্ড়ঢ়য়"), "words": set(["এবং", "একটি", "করুন", "জন্য", "সঙ্গে"])},
+    "br": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसह़ािीुূृेैोौ्"), "words": set(["आरो", "एसे", "मोनसे", "माव", "आव"])},
+    "doi": {"chars": set("अआइईउऊएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूेैोौ्"), "words": set(["ते", "दे", "ऐ", "इक", "ओह्"])},
+    "gu": {"chars": set("અઆઇઈઉઊઋએઐઓઔકખગઘઙચછજઝઞટઠડઢણતથદધનપફબભમયરલવશષસહ઼ાિીુૂૃેૈોૌ્"), "words": set(["અને", "એક", "માટે", "છે", "સાથે"])},
+    "hi": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्"), "words": set(["और", "है", "एक", "में", "के"])},
+    "kn": {"chars": set("ಅಆಇಈಉಊಋಎಏಐಒಓಔಕಖಗಘಙಚಛಜಝಞಟಠಡಢಣತಥದಧನಪಫಬಭಮಯರಲವಶಷಸಹಳಱಾಿೀುೂೃೆೇೈೊೋೌ್"), "words": set(["ಮತ್ತು", "ಒಂದು", "ಹೇಗೆ", "ನಾನು", "ಇದೆ"])},
+    "ks": {"chars": set("اآبپتٹثجچحخدڈذرڑزژسشصضطظعغفقکگلمنوھءییے"), "words": set([" تہٕ", "چھُ", "اکھ", "منز", "کیتھ"])},
+    "kok": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्"), "words": set(["आनी", "एक", "कर", "खातीर", "कडेन"])},
+    "mai": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्"), "words": set(["आ", "एक", "हम", "अछि", "क"])},
+    "ml": {"chars": set("അആഇഈഉഊഋഎഏഐഒഓഔകഖഗഘങചഛജഝഞടഠഡഢണതഥദധനപഫബഭമയരലവശഷസഹളഴറാിീുൂൃെേൈൊോൌ്"), "words": set(["ഒരു", "மற்றும்", "എങ്ങനെ", "ഞാൻ", "ഇതു"])},
+    "mni": {"chars": set("ꯑ꯲꯳꯴꯵꯶꯷꯸꯹꯺꯻꯼꯽꯾꯿ꯀꯂꯃꯄꯅꯆꯇꯈꯉꯊꯋꯌꯍꯎꯏꯐꯑ"), "words": set(["ꯗꯥ", "ꯑꯃꯥ", "ꯀꯔꯤ", "ꯑꯩꯅꯥ", "ꯑꯁꯤ"])},
+    "mr": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझ��टठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्ळ"), "words": set(["आणि", "एक", "आहे", "मी", "तू"])},
+    "ne": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्"), "words": set(["र", "एक", "हो", "म", "तिमी"])},
+    "or": {"chars": set("ଅଆଇଈଉଊଋଏଐଓଔକଖଗଘଙଚଛଜଝଞଟଠଡଢଣତଥଦଧନପଫବଭମଯରଲଳବଶଷସହକ୍ଷାିୀୁୂୃେୈୋୌ୍"), "words": set(["ଏବଂ", "ଗୋଟିଏ", "କରନ୍ତୁ", "ପାଇଁ", "ସହିତ"])},
+    "pa": {"chars": set("ਅਆਇਈਉਊਏਐਓਔਕਖਗਘਙਚਛਜਝਞਟਠਡਢਣਤਥਦਧਨਪਫਬਭਮਯਰਲਵਸ਼ਸਹਖ਼ਗ਼ਜ਼ੜਫ਼ਲ਼ਿੀੁੂੇੈੋੌ੍"), "words": set(["ਅਤੇ", "ਇੱਕ", "ਹੈ", "ਵਿੱਚ", "ਨੂੰ"])},
+    "sa": {"chars": set("अआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलवशषसहािीुूृेैोौ्"), "words": set(["च", "एकः", "अस्ति", "अहम्", "त्वम्"])},
+    "sat": {"chars": set("ᱚᱟᱤᱥᱩᱨᱮႅᱳ鄴ᱠᱜᱝᱪᱡᱧଟଡᱬᱛᱫ Narayan pur pᱷᱵᱶᱷ"), "words": set(["ᱟᱨ", "ᱫᱚ", "হয়", "ఒకటి", "మరియు"])},
+    "sd": {"chars": set("اآبڀتٽثپجڄ جھچحخڌدڏڊذرزڙژسشصضطظعغفڦقڪکگڳڱلمنوھ ءي"), "words": set(["۽", "هڪ", "آهي", "۾", "کي"])},
+    "ta": {"chars": set("அஆஇஈஉஊஎஏஐஒஓஔகஙசஞடணதநனபமயரலவழளஷஸஹாిീுூெேைொோௌ்"), "words": set(["மற்றும்", "ஒரு", "வேண்டும்", "நான்", "இது"])},
+    "te": {"chars": set("అఆఇఈఉఊఋఎఏఐఒఓఔకఖగఘఙచఛజఝఞటఠడఢణతథదధనపఫబభమయరలవశషసహళక్షఱాిీుూృెేైొోౌ్"), "words": set(["మరియు", "ఒక", "வேண்டும்", "నేను", "ఇది"])},
+    "ur": {"chars": set("اآبپتٹثجچحخدڈذرڑزژسشصضطظعغفقکگلمنوھءییے"), "words": set(["اور", "ہے", "ایک", "میں", "کے"])},
 }
 
+LANGUAGE_CODE_TO_NAME = {v: k for k, v in LANGUAGE_DATA.items()}
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# Load Indic Conformer model (assumes custom forward handles decoding strategy)
+# Load Indic Conformer model
+print("Loading IndicConformer model...")
 model = AutoModel.from_pretrained("ai4bharat/indic-conformer-600m-multilingual", trust_remote_code=True).to(device)
 model.eval()
+print("✅ Model loaded successfully.")
+
+
+def identify_language(text: str) -> str | None:
+    """Identifies the language of a given text based on character sets and common words."""
+    if not text.strip():
+        return None
+
+    scores = {lang: 0 for lang in LANGUAGE_DATA}
+    text_chars = set(text)
+    # Use regex to split words, handling various scripts
+    text_words = set(re.split(r'[\s,.:;!?]+', text))
+
+    for lang_code, data in LANGUAGE_DATA.items():
+        char_score = len(text_chars.intersection(data["chars"]))
+        word_score = len(text_words.intersection(data["words"]))
+        
+        # Give more weight to character matches as they are a stronger signal of the script
+        scores[lang_code] = (char_score * 2) + word_score
+
+    # Identify the language with the highest score
+    # Return None if the highest score is very low, indicating a poor match
+    max_score = max(scores.values())
+    if max_score < 3: # Heuristic threshold to prevent misidentification on noise
+        return None
+        
+    identified_code = max(scores, key=scores.get)
+    return identified_code
+
 
 @spaces.GPU
-def transcribe_ctc_and_rnnt(audio_path, language_name):
-    lang_code = LANGUAGE_NAME_TO_CODE[language_name]
+def transcribe_and_identify(audio_path):
+    if not audio_path:
+        return "Please provide an audio file.", "", ""
 
     # Load and preprocess audio
-    waveform, sr = torchaudio.load(audio_path)
-    waveform = waveform.mean(dim=0, keepdim=True) if waveform.shape[0] > 1 else waveform
-    waveform = torchaudio.functional.resample(waveform, sr, 16000).to(device)
+    try:
+        waveform, sr = torchaudio.load(audio_path)
+        waveform = waveform.mean(dim=0, keepdim=True) if waveform.shape[0] > 1 else waveform
+        waveform = torchaudio.functional.resample(waveform, sr, 16000).to(device)
+    except Exception as e:
+        return f"Error loading audio: {e}", "", ""
 
     try:
-        # Assume model's forward method takes waveform, language code, and decoding type
+        # 1. Perform a fast, initial transcription using a pivot language (Hindi)
         with torch.no_grad():
-            transcription_ctc = model(waveform, lang_code, "ctc")
-            transcription_rnnt = model(waveform, lang_code, "rnnt")
+            initial_transcription = model(waveform, "hi", "ctc")
+
+        # 2. Identify the language from the initial transcription
+        identified_lang_code = identify_language(initial_transcription)
+
+        if not identified_lang_code:
+            detected_lang_str = "Language not detected or unsupported."
+            return detected_lang_str, initial_transcription + " (pivot)", "Could not perform final transcription."
+
+        detected_lang_str = f"Detected Language: {LANGUAGE_CODE_TO_NAME.get(identified_lang_code, 'Unknown')}"
+
+        # 3. Perform the final, high-quality transcription using the identified language
+        with torch.no_grad():
+            transcription_ctc = model(waveform, identified_lang_code, "ctc")
+            transcription_rnnt = model(waveform, identified_lang_code, "rnnt")
+
     except Exception as e:
-        return f"Error: {str(e)}", ""
+        return f"Error during transcription: {str(e)}", "", ""
+
+    return detected_lang_str, transcription_ctc.strip(), transcription_rnnt.strip()
 
-    return transcription_ctc.strip(), transcription_rnnt.strip()
 
 # Gradio UI
-with gr.Blocks() as demo:
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
     gr.Markdown(f"## {DESCRIPTION}")
+    gr.Markdown("Upload or record audio in any of the 22 supported Indian languages. The app will automatically detect the language and provide the transcription using both CTC and RNNT decoding.")
+    
     with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=1):
             audio = gr.Audio(label="Upload or Record Audio", type="filepath")
-            lang = gr.Dropdown(
-                label="Select Language",
-                choices=list(LANGUAGE_NAME_TO_CODE.keys()),
-                value="Hindi"
-            )
-            transcribe_btn = gr.Button("Transcribe (CTC + RNNT)")
-        with gr.Column():
-            gr.Markdown("### CTC Transcription")
-            ctc_output = gr.Textbox(lines=3)
-            gr.Markdown("### RNNT Transcription")
-            rnnt_output = gr.Textbox(lines=3)
-
-    transcribe_btn.click(fn=transcribe_ctc_and_rnnt, inputs=[audio, lang], outputs=[ctc_output, rnnt_output], api_name="transcribe")
+            transcribe_btn = gr.Button("Transcribe", variant="primary")
+        
+        with gr.Column(scale=2):
+            detected_lang_output = gr.Label(label="Language Detection Result")
+            gr.Markdown("### RNNT Transcription (More Accurate)")
+            rnnt_output = gr.Textbox(lines=3, label="RNNT Output")
+            gr.Markdown("### CTC Transcription (Faster)")
+            ctc_output = gr.Textbox(lines=3, label="CTC Output")
+
+    transcribe_btn.click(
+        fn=transcribe_and_identify, 
+        inputs=[audio], 
+        outputs=[detected_lang_output, ctc_output, rnnt_output],
+        api_name="transcribe"
+    )
 
 if __name__ == "__main__":
-    demo.queue().launch()
+    demo.queue().launch(share=True)
+
+```