IBM Granite Speech 5.0: Free Open-Source Transcription on Your Mac — Setup + the Subtitle Gig Play

In one sentence: it’s like a personal stenographer that lives inside your Mac — you play it a meeting recording or a video’s audio, and it types out every word it hears, in 6 languages, for free. No subscription, no audio ever leaving your computer.

Two terms, plainly: open source means IBM published the model’s full recipe for free — anyone can download it and use it, including for paid client work. Transcription just means converting spoken audio into written text.

Released today, October 22, 2026, per IBM’s official announcement. IBM says it ranks #2 among open-weight models on the Open ASR Leaderboard — vendor claim, not independently verified. What this means for you: IBM is pitching this free model as near the top of the public accuracy rankings, but nobody outside IBM has confirmed that yet.

Why it matters to you

• Meeting recordings become searchable text. A 1-hour Zoom recording turns into notes you can search instead of re-listening.
• Videos get subtitles. Creators need captions for accessibility and for viewers who watch on mute — this does the heavy lifting for free.
• No monthly fee. Paid transcription tools (Otter, Rev and the like) charge a monthly subscription or per-minute fees — plans change often, so check their sites for current numbers. This one is $0, forever, after setup.
• Private by design. The audio never uploads anywhere, because the model runs on your own machine. What this means for you: sensitive meetings stay on your laptop.

Price: $0

The model is released under the Apache 2.0 license — free to download, free to use, free to build a business on. What this means for you: you can legally charge clients for transcription work done with it and keep 100%.

Two sibling releases dropped the same day, worth one line each: Turbo CTC, an extra-fast lightweight version (6 more languages) for edge devices, and 4.1 2B Plus, which separates multiple speakers — it labels who said what in a group conversation.

The money play

Save: if you currently pay for a transcription subscription, this replaces most of it. What this means for you: the subscription you cancel is pure savings starting month one, for 15 minutes of setup.

Earn — the subtitle / meeting-notes gig:

• Who pays: YouTubers, podcasters, course creators, and small teams sitting on hours of un-captioned video or meeting recordings.
• What you sell: “I turn your audio into clean subtitles or meeting notes, 48-hour turnaround.”
• What to charge: set your own test price — start low for the first 3 clients (a small per-minute-of-audio rate), then raise it once you have samples to show. This is a suggestion to test the market, not a market rate.
• 7-day validation: Days 1–2, set up and transcribe 3 public videos as portfolio samples. Days 3–4, message 20 creators offering one free 5-minute sample. Days 5–7, convert to paid — the goal is 1 paying client or 5 warm leads. If nobody replies after 20 genuine outreaches, drop it: the setup cost you nothing.

Setup in 4 steps (Apple Silicon Mac)

Ready to run it? Here are the 4 steps.

Honest upfront: this involves typing a few commands into the Terminal app. No coding knowledge needed — just copy, paste, and press Enter.

Step 1. Install the runner:

pip install mlx-audio

(If pip isn’t found, install Python 3.10+ from python.org first. You may also need brew install ffmpeg.)

Step 2. Get the model. The official repo is ibm-granite/granite-speech-5.0-2b on Hugging Face (IBM’s own organization, Apache 2.0). The first run downloads it automatically — about 2 billion parameters, a few GB on disk.

Step 3. Transcribe an audio file:

python -m mlx_audio.stt.generate \
  --model ibm-granite/granite-speech-5.0-2b \
  --audio meeting.m4a \
  --output-path ./transcripts \
  --format json

Replace meeting.m4a with your own file. The command shape is verified against the mlx-audio README — but this is a day-one release, so if the runner complains about the model, check the model card on the official repo for the recommended setup. IBM also published a GGUF build (ibm-granite/granite-speech-5.0-2b-GGUF) for the llama.cpp path.

Step 4. Open the ./transcripts folder and grab your text. Done — first result in about 10 minutes.

What comes out looks like this (illustrative sample):

[
  {"start": 12.4, "end": 18.9, "text": "Okay, let's kick off the Q3 planning call."},
  {"start": 19.1, "end": 26.3, "text": "First item: the November launch moves to the 15th."}
]

Each chunk is timestamped, so you can click straight back to the audio — the raw material for subtitles or meeting notes.

Pitfalls and limits

• Apple Silicon Mac only for this path — mlx-audio doesn’t run on Windows or Intel Macs.
• Command line required. There’s no pretty app window; you type commands.
• 6 languages only. Check the model card on the official repo to confirm yours is supported before promising a client.
• Accuracy is IBM’s own claim (see above — vendor claim, not independently verified).
• Day-one rough edges. Brand-new releases sometimes need a week for community tooling to catch up; the GGUF + llama.cpp fallback exists if the main path stumbles.

Free tinkering vs. paid convenience

Want the same “audio becomes text on my Mac” result without the terminal? We covered Koi today — a $59 Mac app that does this point-and-click. Granite is the free DIY version; Koi is the pay-for-convenience version. Pick your trade-off.

Sources

• IBM Granite official announcement, October 22, 2026 — release date, 6-language transcription + translation, Open ASR Leaderboard claim, Turbo CTC and 4.1 2B Plus siblings.
• huggingface.co/ibm-granite — official org page; model repos granite-speech-5.0-2b, granite-speech-5.0-470m-turboctc, granite-speech-4.1-2b-plus, and granite-speech-5.0-2b-GGUF all listed and updated within the last day, verified October 22, 2026.
• mlx-audio GitHub README — STT CLI syntax (python -m mlx_audio.stt.generate), install steps, and Apple Silicon requirements.