faster-whisper Is 8.9X Faster Than Whisper (Same Model, $0)

Show Notes

Watch the video: https://www.youtube.com/watch?v=mgctslLxLH0

13 minutes of audio. OpenAI's own Whisper takes 2m23s. faster-whisper does it in 17s on the same GPU, same accuracy, $0.
Playbooks + my answers, free - Church of SEO Jesus: https://www.skool.com/church-of-seo-jesus

What you get in this video:
- The public benchmark row that proves 8.9x faster on identical hardware
- The exact CPU install that transcribes Kennedy audio in 2.88s on a laptop
- The dollar math for a sales team doing 1,000 hours of calls a month
- Where faster-whisper beats the paid Python reference on Word Error Rate
- The two catches your developer hits on Monday (CUDA pins + speaker labels)

Join the Church of SEO Jesus (free Skool community) for the copy-paste install one-liners, the exact hardware I benched on, and the vendor swap-out spreadsheet operators are running this week: https://www.skool.com/church-of-seo-jesus

Sources:
- faster-whisper repo + benchmark tables: https://github.com/SYSTRAN/faster-whisper
- openai/whisper reference: https://github.com/openai/whisper

Chapters below name every entity so you can jump to the receipt you need.

Chapters:
00:00 faster-whisper vs Whisper API: 17s vs 2m23s
00:23 What faster-whisper actually is (CTranslate2 + Whisper weights)
01:13 The old paid-transcription assumption just died
01:31 RTX 3070 Ti benchmark: openai/whisper vs faster-whisper
01:48 CPU-only run: Intel i7-12700K finishes in 51s
02:13 Free Church install one-liner + subscribe
02:50 Live demo: pip install faster-whisper 1.0.3 + Kennedy audio
03:18 distil-whisper-large-v3: lower WER than transformers SDPA
03:32 Money math: $360/mo Whisper API vs $0 faster-whisper
03:54 Otter, Rev, AssemblyAI: what each one just lost
04:27 Open the GitHub page yourself: verify the row
04:51 5 business use cases that pay back this week
05:40 Two catches: CUDA/cuDNN pins + pyannote for speaker labels
06:28 Why your bookkeeper still sees the line item
06:55 Church of SEO Jesus: install scripts + swap-out sheet
07:15 Next up: GPT-4-class outputs offline on a 2-year-old laptop

Audio edition of the Stewart Vickers AI YouTube video of the same name. Full transcript on this episode page.

What is Stewart Vickers AI?

The audio edition of the Stewart Vickers AI YouTube channel. Each episode takes one AI release - a model, a tool, a price cut - and translates it into money for business owners: what got cheaper, who got undercut, and what to do about it this week. Operator credibility over theory, specific numbers over vibes. Every episode links the original video and the free Church of SEO Jesus community where the prompts and workflows live.

13 minutes of audio, OpenAI's own Whisper takes 2 minutes 23 seconds to transcribe it. A free fork of the same Modal does the identical job in 17 seconds. Same graphics card, same accuracy, zero API calls. If you are still paying Rev, Otter, Descript or the OpenAI Whisper API by the minute, someone on your team owes you a conversation this week. The tool is called Faster Whisper. It is a reimplementation of OpenAI's own Whisper transcription Modal and it runs up to 4 times faster than the original for identical output.

25,000 stars on GitHub, fully open source, $0 a month. It is the reason the entire PR minute transcription market just quietly lost its excuse. Here is the click contract. In the next 6 minutes I will show you the benchmark receipts on the official repo, the exact command that runs on a laptop with no GPU, the dollar math for a sales team doing 1000 hours of calls a month, and the two catches your developer needs to know about before you kill the subscription. And the last one is why most business owners have not switched yet. Start with what changed.

The paid transcription market was built on one assumption. High quality speech to text needs a hosted API because the Modal is too expensive to run locally. Not anymore. Faster Whisper is the same Whisper Modal OpenAI publishes, wrapped in a faster inference engine called cTranslate2. Same weights, faster maths. The benchmark table is public, same 13 minute audio file, same Nvidia graphics card, same accuracy settings, OpenAI's implementation, 2 minutes 23 seconds.

Faster Whisper at its fastest setting, 16 seconds. Almost 9 times faster on the same hardware. And on a plain office laptop CPU with no graphics card at all, it finishes the same file in 51 seconds. 8 times faster than the reference version, no GPU, no cloud. The headline for a business owner is this. Transcription is now a free variable in your cost stack.

Not cheaper, free. Once the Modal is loaded on a machine you already own or rent by the hour, PR minute pricing is a subscription you are choosing to keep paying. If you own a business, subscribe. I cover every one of these open source drops the day it ships, so the next Whisper class release finds you instead of you finding it 6 weeks late on an invoice. The pinned comment under this video has the exact install command the benchmark link and the cost calculator I just walked through. And the Church of SEO Jesus is my free school community linked below, where operators are running these vendor swaps live and I answer in there when your developer setup breaks on Monday.

The install one-liner for both CPU and GPU is in the church post. Now watch the real install. One pip command, a 1500 kilobyte audio file downloaded from OpenAI's own test set. 88 seconds, and returned the sentence word perfect. 8. 8 times faster than the audio it just transcribed, on hardware sitting in every office in the country.

There is more. On the harder reference benchmark, Faster Whisper posts a lower word error rate than the paid Python library the industry treats as default, while finishing in roughly half the wall time. More accurate and faster and free. Now the money math. 36 PR hour comes to $360 a month, $4,320 a year. Same volume on Faster Whisper running on a machine you already have, $0.

OtterPro is around $17 PR user PR month, capped at 1,200 minutes each, which for a 10 person team is $2,000 a year and hard caps your usage the second you get busy. 20 PR minute. 37 an hour. Every one of them just got undercut by the same open source repo, to $0. None of those companies get to charge for the Modal itself anymore. What you pay them for is the browser tab you log into to see the transcript.

Open the Faster Whisper GitHub page yourself right now. Scroll to the benchmark table. Look at the row that says 1 minute 3 seconds versus 2 minutes 23 seconds. That is the proof, not me telling you. Still ahead, the five business use cases where this pays back the same week, and the two catches your developer will hit on Monday morning if you send them at it cold. Here is where operators pay back fastest, in order.

Sales call libraries are the biggest wallet. A team recording 30 calls a week is hitting on hundreds of hours a month of audio that needs to be searchable, and the paid tool bill for that alone justifies a $300 used graphics card in a weekend. Podcast repurposing is second. Four hours of raw episode audio a week becomes free to transcribe, so cutting clips and writing show notes stops being a PR-minute decision. Meeting notes for a small team is third. Customer support voicemail is fourth.

And YouTube captions for your own channel is fifth, and the cheapest one to test tonight. None of that is theory. My phone is full of operators who killed one of these accounts this quarter, in order of size. Two catches your developer needs to know before Monday. First, on a graphics card the install needs the right NVIDIA libraries pinned to a specific CUDA version. Get it wrong and they spend an afternoon on dependency errors instead of transcripts.

Not a hard problem. A known one documented on the repo. Second, speaker labels. Faster Whisper does not tell you who said what. It gives you the words in order. If you need Sarah vs Tom columns on a sales call, you bolt on a second open source tool called PyAnnotode that handles the voices, and your developer wires the two together in a day.

Also known, also solvable. Once it is done, it is done. No renewal, no overage, just words on a page. Neither of these are deal breakers. They are one afternoon of setup. The real question is not whether Faster Whisper works.

The benchmarks say it does. Word error rate lower than the paid Python implementation. Wall time roughly half. The real question is why your bookkeeper is still seeing a monthly line item. For a service, a free repo does faster on hardware you can rent for the price of a coffee. A developer or an agency handles the setup in a day.

The line item disappears the same day. One more thing before you close the tab. The Church of SEO Jesus is my free school community, linked in the description below. And this is where I post the CPU and GPU install one-liners, the exact hardware I benched this on, and the swap-out spreadsheet operators I'm using to cut their transcription line item this week. Free to join. Every viewer of this video has a church post waiting with the copy-paste commands ready to go.

Next up on the channel, the local Modal that just made GPG-4 class outputs run offline on a two-year-old gaming laptop. Same story, different vendor, bigger bill.