Insanely Fast Whisper

Powered by 🤗 Transformers & Optimum

TL;DR - Transcribe 300 minutes (5 hours) of audio in less than 10 minutes - with OpenAI's Whisper Large v2. Blazingly fast transcription is now a reality!⚡️

Basically all you need to do is this:

import torch
from transformers import pipeline

pipe = pipeline("automatic-speech-recognition",
                "openai/whisper-large-v2",
                torch_dtype=torch.float16,
                device="cuda:0")

pipe.model = pipe.model.to_bettertransformer()

outputs = pipe("<FILE_NAME>",
               chunk_length_s=30,
               batch_size=24,
               return_timestamps=True)

outputs["text"]

Not convinced? Here are some benchmarks we ran on a free Google Colab T4 GPU! 👇

🆕 You can now access blazingly fast transcriptions via your terminal! ⚡️

We've added a v1 CLI to enable fast transcriptions. Here's how you can use it.

Install dependencies

pip install -r requirements.txt

Transcribe your audio

python transcribe.py --file_name <filename or URL>

Note: The CLI is opiniated and currently only works for Nvidia GPUs. Make sure to check out the defaults and the list of options you can play around with to maximise your transcription throughput. Run python transcribe.py --help to get all the CLI arguments.

How does this all work?

Here-in, we'll dive into optimisations that can make Whisper faster for fun and profit! Our goal is to be able to transcribe a 2-3 hour long audio in the fastest amount of time possible. We'll start with the most basic usage and work our way up to make it fast!

The only fitting test audio to use for our benchmark would be Lex interviewing Sam Altman. We'll use the audio file corresponding to his podcast. I uploaded it on a wee dataset on the hub here.

Installation

pip install -q --upgrade torch torchvision torchaudio
pip install -q git+https://github.com/huggingface/transformers
pip install -q accelerate optimum
pip install -q ipython-autotime

Let's download the audio file corresponding to the podcast.

wget https://huggingface.co/datasets/reach-vb/random-audios/resolve/main/sam_altman_lex_podcast_367.flac

Base Case

import torch
from transformers import pipeline

pipe = pipeline("automatic-speech-recognition",
                "openai/whisper-large-v2",
                device="cuda:0")

outputs = pipe("sam_altman_lex_podcast_367.flac", 
               chunk_length_s=30,
               return_timestamps=True)

outputs["text"][:200]

Sample output:

We have been a misunderstood and badly mocked org for a long time. When we started, we announced the org at the end of 2015 and said we were going to work on AGI, people thought we were batshit insan

Time to transcribe the entire podcast: 31min 1s

Batching

outputs = pipe("sam_altman_lex_podcast_367.flac", 
               chunk_length_s=30,
               batch_size=8,
               return_timestamps=True)

outputs["text"][:200]

Time to transcribe the entire podcast: 13min 19s

Half-Precision

pipe = pipeline("automatic-speech-recognition",
                "openai/whisper-large-v2",
                torch_dtype=torch.float16,
                device="cuda:0")                

outputs = pipe("sam_altman_lex_podcast_367.flac",
               chunk_length_s=30,
               batch_size=16,
               return_timestamps=True)

outputs["text"][:200]

Time to transcribe the entire podcast: 6min 13s

BetterTransformer w/ Optimum

pipe = pipeline("automatic-speech-recognition",
                "openai/whisper-large-v2",
                torch_dtype=torch.float16,
                device="cuda:0")

pipe.model = pipe.model.to_bettertransformer()

outputs = pipe("sam_altman_lex_podcast_367.flac",
               chunk_length_s=30,
               batch_size=24,
               return_timestamps=True)

outputs["text"][:200]

Time to transcribe the entire podcast: 5min 2s

Roadmap

• Add benchmarks for Whisper.cpp
• Add benchmarks for 4-bit inference
• Add a light CLI script
• Deployment script with Inference API

Community showcase

@ochen1 created a brilliant MVP for a CLI here: https://github.com/ochen1/insanely-fast-whisper-cli (Try it out now!)