simple app to transcribe

.gitignore

@@ -0,0 +1,2 @@
+wandb/
+artifacts/

app.py

@@ -0,0 +1,68 @@
+from datetime import datetime, timedelta
+import os
+
+import gradio as gr
+import torch
+import nemo.collections.asr as nemo_asr
+import wandb
+
+MODEL_HISTORY_DAYS = 180
+WANDB_ENTITY = os.environ.get("WANDB_ENTITY", "tarteel")
+WANDB_PROJECT_NAME = os.environ.get("WANDB_PROJECT_NAME", "nemo-experiments")
+MODEL_NAME = os.environ.get("MODEL_NAME", "CfCtcLg-SpeUni1024-DI-EATLDN-CA:v0")
+
+
+def get_device():
+    if torch.cuda.is_available():
+        return "cuda"
+    else:
+        return "cpu"
+
+
+# run = wandb.init(entity=WANDB_ENTITY, project=WANDB_PROJECT_NAME)
+wandb_api = wandb.Api(overrides={"entity": WANDB_ENTITY})
+
+artifact = wandb_api.artifact(f"{WANDB_ENTITY}/{WANDB_PROJECT_NAME}/{MODEL_NAME}")
+artifact_dir = artifact.download()
+
+# find the model (ending with .nemo) in the artifact directory
+model_path = [
+    os.path.join(root, file)
+    for root, dirs, files in os.walk(artifact_dir)
+    for file in files
+    if file.endswith(".nemo")
+][0]
+
+model = nemo_asr.models.EncDecCTCModelBPE.restore_from(
+    model_path, map_location=get_device()
+)
+
+
+def transcribe(audio_file):
+    transcription_file = model.transcribe([audio_file])[0]
+    return transcription_file
+
+
+demo = gr.Blocks()
+
+with demo:
+    gr.Markdown(
+        """
+        # ﷽
+        """
+    )
+    with gr.Row():
+        audio_file = gr.Audio(source="upload", type="filepath", label="File")
+
+    with gr.Row():
+        output_file = gr.TextArea(label="Audio Transcription")
+
+    b1 = gr.Button("Transcribe")
+
+    b1.click(
+        transcribe,
+        inputs=[audio_file],
+        outputs=[output_file],
+    )
+
+demo.launch()

packages.txt

@@ -0,0 +1,2 @@
+ffmpeg
+libsndfile1

requirements.txt

@@ -0,0 +1,4 @@
+torchaudio
+nemo_toolkit[asr]==1.19.0
+hydra-core
+wandb