# AI use cases for Whisper CPP (ggml-org/whisper.cpp) Repo: [ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp) | Use cases: 3 | Page: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/ whisper.cpp is a plain C/C++ port of OpenAI's Whisper speech recognition model with no dependencies. It runs fully offline on CPUs, Apple Silicon (Metal and Core ML), NVIDIA, AMD and Vulkan GPUs, supports integer quantization and voice activity detection, and ships a whisper-cli tool plus examples for iOS, Android, WebAssembly and Raspberry Pi. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Transcribe audio files offline on your own machine: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/transcribe-audio-offline-with-whisper-cpp/ - Run bigger Whisper models on small hardware with quantization: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/shrink-whisper-models-with-quantization/ - Speed up transcription on a Mac with the Neural Engine: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/speed-up-whisper-on-apple-silicon-with-core-ml/ ## How to transcribe audio files offline on your own machine with Whisper CPP Repo: [ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp) | Difficulty: Beginner | Time: 20 minutes | Basis: From the README Build whisper-cli, download a ggml model and turn recordings into text without any cloud service. **Who it is for:** A journalist or podcaster with hours of interviews who cannot upload the audio anywhere. **The problem:** Cloud transcription costs money per minute and means handing private recordings to a third party. **The outcome:** A local transcript of each recording, produced entirely on your own hardware. ### Steps 1. Run `git clone https://github.com/ggml-org/whisper.cpp.git` and cd into the folder 2. Download a model with `sh ./models/download-ggml-model.sh base.en` 3. Build the tool: run `cmake -B build` and then `cmake --build build -j --config Release` 4. Convert your recording to 16-bit WAV with `ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav` 5. Transcribe it with `./build/bin/whisper-cli -f samples/jfk.wav`, swapping in your own file, and check `./build/bin/whisper-cli -h` for output options ### Prompt to paste into your agent ```text Clone ggml-org/whisper.cpp, build whisper-cli with CMake, download the base.en model, convert every MP3 in to 16 kHz mono WAV with ffmpeg and transcribe each one. Save one text file per recording next to the audio. ``` ### Caveat whisper-cli only reads 16-bit WAV, so every other format needs an ffmpeg conversion first, and larger models need several GB of RAM. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/transcribe-audio-offline-with-whisper-cpp/ ## How to run bigger Whisper models on small hardware with quantization with Whisper CPP Repo: [ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Quantize a ggml Whisper model to Q5_0 so it uses less memory and disk with the same commands. **Who it is for:** Someone transcribing on a Raspberry Pi, an old laptop or a small server where the full model barely fits. **The problem:** The more accurate Whisper models need gigabytes of memory, which rules them out on modest machines. **The outcome:** A quantized model file that whisper-cli loads like any other, using less memory and disk space. ### Steps 1. Build Whisper CPP with `cmake -B build` and `cmake --build build -j --config Release` 2. Download the model you want to shrink, for example with `sh ./models/download-ggml-model.sh base.en` 3. Quantize it: `./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0` 4. Run the quantized model: `./build/bin/whisper-cli -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav` 5. Compare speed and accuracy against the original on a few of your own files before switching ### Prompt to paste into your agent ```text Using ggml-org/whisper.cpp, quantize the ggml model to q5_0 with the quantize tool, then transcribe with both the original and the quantized model. Report memory use, run time and any differences in the transcripts. ``` ### Caveat Quantization trades a little accuracy for size, so check transcripts on your own audio before relying on it. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/shrink-whisper-models-with-quantization/ ## How to speed up transcription on a Mac with the Neural Engine with Whisper CPP Repo: [ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README Generate a Core ML encoder so Whisper CPP runs the encoder on the Apple Neural Engine, more than 3x faster than CPU only. **Who it is for:** A Mac user who transcribes often and wants the fastest local setup the hardware allows. **The problem:** CPU only transcription on a laptop is slow for long recordings and keeps the fans spinning. **The outcome:** whisper-cli loading a Core ML encoder and transcribing noticeably faster on Apple Silicon. ### Steps 1. Install the Python packages: `pip install ane_transformers`, `pip install openai-whisper` and `pip install coremltools` 2. Make sure Xcode is installed and run `xcode-select --install` for the command line tools 3. Generate the encoder with `./models/generate-coreml-model.sh base.en`, which creates models/ggml-base.en-encoder.mlmodelc 4. Build with Core ML: `cmake -B build -DWHISPER_COREML=1` then `cmake --build build -j --config Release` 5. Run `./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav` and look for the Core ML model loaded line ### Prompt to paste into your agent ```text On my Apple Silicon Mac, set up ggml-org/whisper.cpp with Core ML support: install the Python requirements, generate the Core ML encoder for base.en, rebuild with WHISPER_COREML=1 and transcribe samples/jfk.wav. Tell me how long it took compared with the CPU build. ``` ### Caveat The first run on a device is slow while the Neural Engine compiles the model, and the README recommends macOS Sonoma or newer to avoid transcription hallucination. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/ggml-org-whisper-cpp/speed-up-whisper-on-apple-silicon-with-core-ml/