Working voice pipeline: Bangla STT and TTS in Streamlit app

.gitignore

@@ -0,0 +1,3 @@
+venv/
+__pycache__/
+*.pyc

app.py

@@ -0,0 +1,21 @@
+import streamlit as st
+from voice_functions import listen, speak
+import subprocess
+
+st.title("KrishiKotha - কৃষি সহায়ক")
+
+audio_file = st.file_uploader("আপনার প্রশ্ন রেকর্ড করে আপলোড করুন", type=["wav", "mp3", "m4a"])
+
+if audio_file is not None:
+    with open("temp_input.wav", "wb") as f:
+        f.write(audio_file.getbuffer())
+
+    # Convert to proper format using ffmpeg
+    subprocess.run(["ffmpeg", "-y", "-i", "temp_input.wav", "-ar", "16000", "-ac", "1", "temp_fixed.wav"])
+
+    text = listen("temp_fixed.wav")
+    st.write("**আপনি বলেছেন:**", text)
+
+    # For now, just echo back — later this becomes Pair 2's grounded answer
+    speak(text, "response.mp3")
+    st.audio("response.mp3")

test_roundtrip.py

@@ -0,0 +1,7 @@
+from voice_functions import listen, speak
+
+text = listen("test_audio_fixed.wav")
+print("Heard:", text)
+
+speak(text, "roundtrip_output.mp3")
+print("Spoken back — check roundtrip_output.mp3")

test_tts.py

@@ -0,0 +1,10 @@
+from gtts import gTTS
+
+def speak(text: str, output_path: str = "output.mp3") -> str:
+    tts = gTTS(text=text, lang="bn")
+    tts.save(output_path)
+    return output_path
+
+if __name__ == "__main__":
+    speak("আপনার ধান গাছের পাতা হলুদ হয়ে যাচ্ছে, একজন কৃষি বিশেষজ্ঞের পরামর্শ নিন।")
+    print("Saved to output.mp3")

test_whisper.py

@@ -0,0 +1,15 @@
+from transformers import pipeline
+
+# Bangla fine-tuned Whisper model
+asr = pipeline(
+    "automatic-speech-recognition",
+    model="bangla-speech-processing/BanglaASR"
+)
+
+def listen(audio_file_path: str) -> str:
+    result = asr(audio_file_path)
+    return result["text"]
+
+if __name__ == "__main__":
+    result = listen("test_audio_fixed.wav")
+    print("Transcribed text:", result)

voice_functions.py

@@ -0,0 +1,18 @@
+from transformers import pipeline
+from gtts import gTTS
+
+# Speech-to-text (Bangla-specific model)
+asr = pipeline(
+    "automatic-speech-recognition",
+    model="bangla-speech-processing/BanglaASR"
+)
+
+def listen(audio_file_path: str) -> str:
+    result = asr(audio_file_path)
+    return result["text"]
+
+# Text-to-speech
+def speak(text: str, output_path: str = "output.mp3") -> str:
+    tts = gTTS(text=text, lang="bn")
+    tts.save(output_path)
+    return output_path