How to install and use Whisper offline (no internet required)

[nicholasgcotton]

This guide can also be found at Whisper Full (& Offline) Install Process for Windows 10/11

Whisper Full (& Offline) Install Process for Windows 10/11

Purpose: These instructions cover the steps not explicitly set out on the main Whisper page, e.g. for those who have never used python code/apps before and do not have the prerequisite software already installed.

Requirements:

• Full admin rights on your computer.
• A PC with a CUDA-capable dedicated GPU with at least 4GB of VRAM (but more VRAM is better). See: Available models and languages
• For online installation: An Internet connection for the initial download and setup.
• For offline installation: Download on another computer and then install manually using the "OPTIONAL/OFFLINE" instructions below.

Installation

Step 1: Unlisted Pre-Requisites

• Before you can run whisper you must download and install the follopwing items. (For offline installation just download the files on another machine and move them to your offline machine to install them.)
  • NVIDIA CUDA drivers: https://developer.nvidia.com/cuda-downloads
  • Python 3.9 or 3.10 (x64 version) from https://www.python.org/ (Whisper claims to run with >3.7 but as of 2023-01-18 some dependencies require >3.7 but <3.11).
  • FFMPEG
    • To install via Scoop (https://scoop.sh/), in powershell run
      • Set-ExecutionPolicy RemoteSigned -Scope CurrentUser
      • irm get.scoop.sh | iex
      • scoop install ffmpeg
    • OPTIONAL/OFFLINE: Follow instructions here: How to install and use FFMPEG and make sure not to skip the part about adding FFMPEG to the Windows PATH variable.
  • Git for windows from https://gitforwindows.org/
• Reboot after installing these items.

Step 2B: Whipser Install (Online Install for Online Use)

• Open a command prompt and type this command:
  • pip install git+https://github.com/openai/whisper.git
• You may now use Whisper online and no further steps are required.

Step 2B: Whipser Install (Online Install for later Offline Use)

• Open a command prompt and type these commands:
  • pip install git+https://github.com/openai/whisper.git
  • pip install blobfile
• Continue to Step 3: Download Other Required Files

Step 2C: Whipser Install (Offline Install for later Offline Use)

• Option 1: Get the most up to date version of Whisper:
  • Install Python and Git from Step 1 on an second computer you can connect to the internet and reboot to ensure both are working.
  • On the ONLINE machine open a command prompt in any empty folder and type the following commands:
    • pip download git+https://github.com/openai/whisper.git
    • pip download blobfile
• Option 2: Download all the necessary files from here OPENAI-Whisper-20230314 Offline Install Package
• Copy the files to your OFFLINE machine and open a command prompt in that folder where you put the files, and run
  • pip install openai-whisper-20230314.zip (note the date may have changed if you used Option 1 above).
  • pip install blobfile-2.0.2-py3-none-any.whl. (note the version may have changed if you used Option 1 above).
• Continue to Step 3: Download Other Required Files

Step 3: Download Other Required Files (for Offline Use)

• Download Whisper's Language Model files place them in C:\Users[Username]\.cache\whisper Note: If the links are dead updated links can be found at lines 17-27 here: init.py
  • Tiny.En
  • Tiny
  • Base.En
  • Base
  • Small.En
  • Small
  • Medium.En
  • Medium
  • Large-v1
  • Large-v2 (Annoucing the large-v2 model)
• Download Whisper's vocabulary and encoder files. (Per issue 1399).
  • Download Vocab.bpe
  • Download Encoder.json
  • Install the files to a folder of your choosing, e.g. C:\Users[Username]\.cache\whisper.
  • Update file links in your local copy of openai_public.py which will be installed in your python folder e.g. C:\Users[UserName]\AppData\Local\Programs\Python\Python310-32\Lib\site-packagespython3.9/site-packages/tiktoken_ext/openai_public.py to point to where you downloaded the files.
    • Remove the URL "https://openaipublic.blob.core.windows.net/gpt-2/encodings/main/" and replace it with your local copy, e.g. "C:/Users/[Username]/.cache/whisper/vocab.bpe" and "C:/Users/[Username]/.cache/whisper/encoder.json"

def gpt2():
    mergeable_ranks = data_gym_to_mergeable_bpe_ranks(
        vocab_bpe_file="C:/Users/nic/.cache/whisper/vocab.bpe",
        encoder_json_file="C:/Users/nic/.cache/whisper/encoder.json",
    )

Alternative Offline Method

See the pre-compiled .exe version of Whisper provided here: Purfview / Whisper Standalone

[Purfview]

Or just download executable from there: https://github.com/Purfview/whisper-standalone-win

[nicholasgcotton]

True and I do recommend that project, but I see this question asked a lot so I figured it would be useful for some to have the proof of concept that it can be done. (Also not that it matters, but when I started figuring this out that other project wasn't an option yet.)

[64jcl]

Looks like a command line executable. Won't that reload the whole model on every run and be highly inefficient?

[Purfview]

@64jcl

Do you mean if you want to run on the multiple files?
In Discussions there are bat [aka batch] examples how to run all files in one go: Purfview/whisper-standalone-win#29

Executable for Faster-Whisper has implementations to do that out of the box.

[phineas-pta]

it's a very detailed instruction, but i would advise against the files transfer between pc for offline installation, it's quite tricky

[nicholasgcotton]

Oh I agree it's very annoying, but it's unavoidable if you want to set the system up on an offline machine. I had to do this due to security rules at my org, and I figured I might as well share the steps once I'd worked it out.

[fromoven]

Hi, I've run into problems installing for offline use because of organisation policy. I've followed the steps and it craps at the openai-whisper install with the same error: installing build dependencies... pip subprocess... NewConnectionError... pip._vendor.urlib3... /simple/setuptools... couldn't find setuptools>=40.8.0. Setuptools is installed with latest version; I tried updating PIP - use PIP3 instead - specify "setuptools==68.0.0" in the install command. Same error. I even tried another computer, re-downloaded the packages with the same result (both times I had Python3.10. I've look at the forum and see that there's someone else who had the same issue (#1556) and a reply points back to the steps here. Any ideas what might be the issue ? Thanks

[LqdtrX]

fromoven, got the same error. Solved:

1. Download requirements.txt from https://github.com/openai/whisper.git and put it along with other files
2. Run the command: pip install --no-index --find-links ./ -r requirements.txt
3. Unpack openai-whisper-20231106.zip and run the command from the corresponding folder: python setup.py install

[IsaacKim9876]

So by running Whisper api on our own pcs we dont have to pay openAI to use it?

[Purfview]

Yeap, it's free.

[palladinoandrea]

but also on 2024?

[docpainting]

Why

[EnnuiThoughts]

pip install blobfile-2.0.2-py3-none-any.whl

this returns a no such file exist error

what should i do?

[RedBluePrinter]

Check if the files version changed currently blobfile-2.0.2-py3-none-any.whl -> blobfile-2.1.1-py3-none-any.whl

[imious]

hi i got this problem with whisper-faster standalone:

[Purfview]

Read there: Purfview/whisper-standalone-win#104

[qingtian305]

I like this project very much. It works very well. Does whisper-faster have a method for calling programs? I want the engine to be on standby for inference all the time after loading, instead of having to reason from the command line every time, which is slow.

[qingtian305]

Also, I wonder if this project can be accelerated with openvino?

[Purfview]

I want the engine to be on standby...

I don't think that's possible with command line executables, it quits when it finishes work on all files.

[gleithoernchen]

I also installed it on a dedicated server and was wondering if there is another way to use it than with the cmd. Is there an api I can call? This would be easier if I want to use it from other servers in this env.

[Baka-14]

Did you find any api to use it ?

[64jcl]

I personally resorted to NodeJS implementation called @xenova/transformers. It is quite easy to use and you just load the model once and can then call on it multiple times. It is easy to rewrite as a server api then:

https://huggingface.co/docs/transformers.js/guides/node-audio-processing

[StepHaze]

Is there an instruction on how to install and use Whisper offline (no internet required) for Ubuntu?

[Purfview]

I made Whisper executable only for Windows but there is Faster-Whisper executable for Linux:
https://github.com/Purfview/whisper-standalone-win/releases

It mirrors functionality of Whisper when it's much faster.

[Sethecientos]

I have a problem with large and large-v2 models.

I can use offline everything but them. When I try to select them, whisper tries to connect to the internet to check something, but only with those two models. I redownloaded few times and still happens.

Any idea?

[Purfview]

Do you use the latest version of Whisper?

[Sethecientos]

Yes, last version, now I could load large-v2 in python, but in the command line using whisper audio.mp --model large-v2, sometimes fails, sometimes doesn't, large failed everytime.

GPU is not a issue.

[Purfview]

Strange, what if you use "exe" from there? -> https://github.com/Purfview/whisper-standalone-win/releases

[misutoneko]

@Sethecientos it might be a good idea to check the checksums of your local model files manually, just in case they've been corrupted somehow (it could even be a hardware problem, like hd dying). The URLs listed here contain the sha256sums for each file.

[tyomart]

Hi! You can hate me, but why WhiPser guys? Is there something we should not know?