🚧 transcribe_audio_with_whisper

app.py

@@ -17,6 +17,7 @@ from tools import (
     count_tables_in_wikipedia_section,
     extract_nth_table_in_wikipedia_section,
     wikipedia_featured_articles_title,
+    transcribe_audio_with_whisper,
 )
 
 # (Keep Constants as is)
@@ -74,6 +75,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
                 count_tables_in_wikipedia_section,
                 extract_nth_table_in_wikipedia_section,
                 wikipedia_featured_articles_title,
+                transcribe_audio_with_whisper,
             ],
             verbosity_level=2,
             additional_authorized_imports=authorized_imports,

requirements.txt

@@ -2,4 +2,6 @@ gradio
 pandas
 requests
 smolagents[toolkit]
+torch
+transformers
 wikipedia-api

tools.py

@@ -2,6 +2,9 @@ from bs4 import BeautifulSoup
 import requests
 import re
 from smolagents import tool
+import torch
+import spaces
+from transformers import pipeline
 
 
 @tool
@@ -177,3 +180,34 @@ def wikipedia_featured_articles_title(year: int) -> str:
              "Wikipedia:Featured articles promoted in 2021".
     """
     return f"Wikipedia:Featured articles promoted in {year}"
+
+
+device = (
+    torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+)
+
+asr_pipeline = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-large-v3",
+    device=device,
+    chunk_length_s=30,
+    return_timestamps=False,
+    generate_kwargs={"task": "transcribe"},
+)
+
+
+@tool
+@spaces.GPU
+def transcribe_audio_with_whisper(filename: str) -> str:
+    """
+    Transcribe an audio file into text using Whisper.
+
+    Args:
+        filename (str): Path to the audio file on disk (e.g., "input/Strawberry pie.mp3").
+
+    Returns:
+        str: Transcribed text of the audio content.
+    """
+    with open(filename, "rb") as f:
+        audio_bytes = f.read()
+    return asr_pipeline(audio_bytes)["text"]