Whisper

You can use Whisper models offline for speech recognition. They work entirely on your local machine, supporting both English-only and multilingual variants. The module supports long audio files and user prompts.

Does the package include the Whisper weights?

Not inside the pip package. pip install intelli[offline] installs the code only, which keeps the package small. You have two ways to get the weights:

• Download on first use. The first time you load a model, its weights download from Kaggle and are cached on your machine (~/.cache/kagglehub). After that, transcription runs offline with no API key.
• Bundle the weights with your app. Download them once into your project folder and ship that folder with your code or Docker image. The machine that runs the app never needs internet. See Bundle the weights with your app.

Available Whisper Models

Model Name	Parameters
whisper_tiny_en	37.18M
whisper_tiny_multi	37.76M
whisper_base_en	124.44M
whisper_base_multi	72.59M
whisper_small_en	241.73M
whisper_small_multi	241.73M
whisper_medium_en	763.86M
whisper_medium_multi	763.86M
whisper_large_multi	1.54B
whisper_large_multi_v2	1.54B

Setup

1. Create a Kaggle account (if you don't already have one) to download any large Whisper model for the first time.
2. Accept the respective license on Kaggle for the chosen model preset.
3. Generate an access token in your Kaggle settings.

Installation

Make sure to install the necessary packages:

pip install --upgrade intelli["offline"]

Importing

# audio
import os
import soundfile as sf
# inference wrapper
from intelli.wrappers.keras_wrapper import KerasWrapper

Using the Model

Load the audio

test_file = "long_audio.ogg"
audio_data, sample_rate = sf.read(test_file)

Run the Inference

wrapper = KerasWrapper(model_name="whisper_large_multi_v2")

result = wrapper.transcript(
    audio_data=audio_data,
    sample_rate=sample_rate,            # e.g., 16000
    language="<|en|>",                  # Use "<|en|>" or another language token
    user_prompt="You are a helpful assistant transcribing teacher's notes.",
    condition_on_previous_text=True,    # Provide context from previous segments
    max_steps=100,                      # Max decoding steps per chunk
    max_chunk_sec=30                    # Max chunk length in seconds
)

Bundle the weights with your app

Use this when the app runs where there is no internet, such as an air-gapped server, a factory PC or a container without network access. You download the weights once on a connected machine, keep them in a folder next to your code, and point the app at that folder.

The folder location comes from the KAGGLEHUB_CACHE environment variable. Without it, the weights go to ~/.cache/kagglehub.

1. Download the weights into your project. Run this once on a machine with internet, from your project folder:

KAGGLEHUB_CACHE=./whisper_weights python -c "from intelli.wrappers.keras_wrapper import KerasWrapper; KerasWrapper(model_name='whisper_tiny_en')"

The whisper_weights folder now holds the model. For whisper_tiny_en it is about 150 MB. Larger models take more space.

2. Point the app at the folder. Set the variable before you load the model:

import os
os.environ["KAGGLEHUB_CACHE"] = "./whisper_weights"

from intelli.wrappers.keras_wrapper import KerasWrapper
wrapper = KerasWrapper(model_name="whisper_tiny_en")

The model now loads from whisper_weights and makes no network call. Copy the folder with your code to any machine, or bake it into a Docker image:

COPY whisper_weights /app/whisper_weights
ENV KAGGLEHUB_CACHE=/app/whisper_weights

Already downloaded a model the default way? Copy ~/.cache/kagglehub into your project instead of downloading again, and set KAGGLEHUB_CACHE to the copied folder.