This is the setup behind the Xentir reel. It uses faster-whisper with the English base.en model on a CPU. The model downloads during the first run; subsequent transcription can use the cached model without sending audio to a transcription API.
You need Python 3.9 or later, a compatible computer, and an audio file you have permission to process. Software and model downloads need internet. Your computer, storage, and electricity still have costs.
Set up
Save the Python script shown below as transcribe.py in a new project folder. Then open a terminal in that folder and run:
python -m venv .venv
On Windows, use the environment's Python directly:
.venv\Scripts\python -m pip install faster-whisper
.venv\Scripts\python transcribe.py "voice-note.wav"
On macOS or Linux:
.venv/bin/python -m pip install faster-whisper
.venv/bin/python transcribe.py "voice-note.wav"
Save the following as transcribe.py beside your audio file:
import argparse
from pathlib import Path
from faster_whisper import WhisperModel
parser = argparse.ArgumentParser()
parser.add_argument("audio", type=Path)
args = parser.parse_args()
if not args.audio.is_file():
parser.error("Audio file not found. Check its filename and folder.")
model = WhisperModel("base.en", device="cpu", compute_type="int8")
segments, info = model.transcribe(str(args.audio), language="en", beam_size=3)
text = " ".join(segment.text.strip() for segment in segments)
output = args.audio.with_name(args.audio.name + ".transcript.txt")
with output.open("x", encoding="utf-8") as handle:
handle.write(text + "\n")
print(text)
print(f"Saved to {output}")
The script refuses to overwrite an existing transcript. Rename or move that transcript before running it again.
What our test actually showed
We used synthetic test audio saying: “Meeting moved to Thursday at three. Send the draft before lunch.” The output was: “Meeting moved to Thursday at 3. Send the draft before lunch.”
That 4.275-second clip took 3.54 seconds to transcribe in our VPS test. This is one short test, not a speed guarantee for your hardware or evidence of accuracy on longer recordings. Model loading, audio length, noise, accents, and hardware affect results.
Check the result
Listen again and check names, numbers, dates, and anything important before using the transcript. English-only base.en is the model demonstrated here. For other languages, consult the project's multilingual model instructions. A model's confidence is not proof that the words are correct.
Follow @xentirmedia for more practical AI demonstrations. No follow confirmation or email signup is required to read this guide.
Source and maintained installation instructions: SYSTRAN/faster-whisper.
