Skip to main content
The Transcription component uses Faster Whisper to convert audio files to timestamped text segments with automatic GPU acceleration.

​
Functions

​
transcribeAudio

Transcribes an audio file into timestamped text segments using the Faster Whisper model. 
transcribeAudio(audio_path: str) -> list[list[str, float, float]]
​
audio_path
str
required
Path to the audio file to transcribe (supports WAV, MP3, and other common formats)
​
transcriptions
list[list]
List of transcription segments, where each segment is [text, start_time, end_time]:
  • text (str): The transcribed text for this segment
  • start_time (float): Start time in seconds
  • end_time (float): End time in seconds

​
Model Configuration

  • Model: base.en (English-only for faster processing)
  • Device: Automatically selects CUDA if available, otherwise CPU
  • Beam Size: 5 (for better accuracy)
  • Language: English (en)
  • Max New Tokens: 128
  • Condition on Previous Text: False (prevents context accumulation)

​
Features

  • Automatic GPU Detection: Uses CUDA if available for 5-10x faster transcription
  • Timestamped Segments: Returns precise start/end times for each phrase
  • Error Handling: Returns empty list on failure with error message
  • Progress Indication: Prints device type and completion status
from Components.Transcription import transcribeAudio

# Transcribe audio file
transcriptions = transcribeAudio("audio.wav")

for text, start, end in transcriptions:
    print(f"[{start:.2f}s - {end:.2f}s]: {text}")

# Output:
# [0.00s - 2.50s]: Hello and welcome to this video
# [2.50s - 5.20s]: Today we'll be discussing AI
# [5.20s - 8.40s]: and its applications in video processing

​
Output Format

Each transcription segment is a list with three elements: 
[
    ["Hello and welcome to this video", 0.0, 2.5],
    ["Today we'll be discussing AI", 2.5, 5.2],
    ["and its applications", 5.2, 8.4]
]
The function automatically detects and uses GPU acceleration if CUDA is available. On CPU, transcription may be significantly slower for long audio files.

​
Performance

  • GPU (CUDA): ~1-2 minutes for a 10-minute audio file
  • CPU: ~5-10 minutes for a 10-minute audio file
  • Model Size: ~150MB download on first run

​
Device Detection

The component prints the detected device on execution: 
print("Transcribing audio...")
Device = "cuda" if torch.cuda.is_available() else "cpu"
print(Device)  # Prints: "cuda" or "cpu"

​
Error Handling

Returns an empty list if transcription fails: 
transcriptions = transcribeAudio("audio.wav")

if len(transcriptions) == 0:
    print("Transcription failed or no speech detected")
else:
    print(f"Successfully transcribed {len(transcriptions)} segments")
Requires faster-whisper and torch packages. For GPU acceleration, CUDA toolkit must be installed. The Whisper model is downloaded automatically on first run (~150MB).

​
Model Details

  • Model Name: base.en
  • Language Support: English only (faster than multilingual models)
  • Accuracy: Good for clear speech, may struggle with heavy accents or background noise
  • Beam Search: Uses beam size of 5 for better accuracy vs. speed tradeoff

Build docs developers (and LLMs) love

Get started for freeTalk to us