Create app.py

app.py

@@ -0,0 +1,84 @@
+import streamlit as st
+import torch
+import librosa
+import soundfile
+import nemo.collections.asr as nemo_asr
+import tempfile
+import os
+import uuid
+from pydub import AudioSegment
+import numpy as np
+import io
+
+SAMPLE_RATE = 16000
+
+# Load pre-trained model
+model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained("stt_en_conformer_transducer_large")
+model.change_decoding_strategy(None)
+model.eval()
+
+
+def process_audio_data(audio_data):
+    # Convert stereo to mono
+    if audio_data.channels == 2:
+        audio_data = audio_data.set_channels(1)
+
+    # Convert pydub audio segment to numpy array
+    audio_np = np.array(audio_data.get_array_of_samples())
+
+    # Resample if necessary
+    if audio_data.frame_rate != SAMPLE_RATE:
+        audio_np = librosa.resample(audio_np, audio_data.frame_rate, SAMPLE_RATE)
+
+    return audio_np
+
+
+def transcribe(audio_np):
+    with tempfile.TemporaryDirectory() as tmpdir:
+        # Save audio data to a temporary WAV file
+        audio_path = os.path.join(tmpdir, f'audio_{uuid.uuid4()}.wav')
+        soundfile.write(audio_path, audio_np, SAMPLE_RATE)
+
+        # Transcribe audio
+        transcriptions = model.transcribe([audio_path])
+
+        # Extract best hypothesis if transcriptions form a tuple (from RNNT)
+        if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+            transcriptions = transcriptions[0]
+
+    return transcriptions[0]
+
+
+st.title("Speech Recognition with NeMo Conformer Transducer Large - English")
+
+# Record audio
+st.write("Click the button below to start recording.")
+record_state = st.checkbox("Recording")
+
+if record_state:
+    # Start recording audio
+    recording = st.audio("", format="audio/wav")
+
+    # Stop recording when checkbox is unchecked
+    recording_file = tempfile.NamedTemporaryFile(delete=False)
+    with recording_file as f:
+        while record_state:
+            audio_data = st.audio_recorder(
+                sample_rate=SAMPLE_RATE,
+                format="wav",
+                data_format="audio/wav"
+            )
+            f.write(audio_data.getvalue())
+
+            # Update recording display
+            audio_data = AudioSegment.from_wav(io.BytesIO(audio_data.getvalue()))
+            recording.audio(audio_data, format="audio/wav")
+            record_state = st.checkbox("Recording")
+
+    # Process and transcribe recorded audio
+    recording_file.seek(0)
+    audio_np = process_audio_data(AudioSegment.from_file(recording_file.name))
+    transcript = transcribe(audio_np)
+
+    st.write("Transcription:")
+    st.write(transcript)