Whisper: transcribe French audio offline
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
You want to transcribe French phone meetings, voice messages, or classes without an online service (software hosted on a subscription, often called SaaS, for Software as a Service)? Whisper is an open-source speech-recognition model released by OpenAI. The openai-whisper package and whisper.cpp run on your machine: the audio does not go to OpenAI if you are not using their cloud API. Only the first download of the model weights may use the internet. This guide extends first steps with Whisper with a concrete chain (ffmpeg plus an MP3 test).
Which tool? Beginner:
pip install openai-whisperplus ffmpeg (steps 1–4). No Python, or a modest PC: whisper.cpp (the alternative section). A graphical interface: LM Studio for one file you drag and drop. That is handy for a one-off test, and a poor fit for multi-hour meetings or a batch of many recordings (there is no long queue the way there is on the command line).
Jobs around the house
| Situation | Why local helps |
|---|---|
| Meeting notes | The recording does not sit on a US server |
| Notes for a relative who is hard of hearing | Editable text, kept in your own archive |
| A chain with an LLM | Transcription, then a summary through Ollama |
| Several files a week | Move to the medium model, or the Go further block, after the MP3 test |
For recordings that include other people, cross-check privacy when data stays fully local.
Hardware you need
| Profile | CPU | GPU | Guidance |
|---|---|---|---|
| Occasional, short files | 4 recent cores | Optional | The small model |
| One-hour meetings | 8 cores | Speeds this up a lot | medium |
| Highest quality | A workstation | 8 GB or more of VRAM (VRAM for image models and LLMs) | large (files longer than about 30 minutes; below that, medium is often safer on 8 GB) |
On CPU alone, plan on about 2× to 10× the audio length, depending on the file. On a GPU with 8 GB of VRAM, save medium or large for files past about 30 minutes of audio. Below that, small keeps memory pressure down and leaves room if Ollama is running at the same time. A decent USB microphone improves the source (optional: the built-in mic is enough for a first test — see first steps with Whisper for the mic choice).
Step 1 — Install ffmpeg and check PATH
ffmpeg is the tool that reads almost every audio format. Whisper relies on it in most installs. Before pip install whisper, confirm ffmpeg. Without this step, step 4 often fails with “ffmpeg not found”.
Windows
- Download an ffmpeg build from the official site or a trustworthy community pack.
- Add the
binfolder to the system PATH. - Step 1b — a required test:
ffmpeg -version
The first line should show ffmpeg version followed by a number. If you see “command not found”, fix PATH before you continue.
Linux (Debian and Ubuntu)
sudo apt update && sudo apt install -y ffmpeg
ffmpeg -version
macOS (Homebrew)
brew install ffmpeg
ffmpeg -version
Step 2 — Install Whisper (Python)
Python 3.10 or newer is a good environment:
python3 -m pip install -U openai-whisper
On some machines, also install PyTorch following the PyTorch docs (CPU, or CUDA for NVIDIA).
An alternative without Python: whisper.cpp
If you do not want Python, or the machine is very limited, whisper.cpp is enough for a WAV or a converted MP3. Compiling it and picking the binary still deserve a real walkthrough.
→ Step 3A in the first-steps Whisper guide: the release, the ggml-small.bin download, the -otxt command, and Windows or Linux fixes.
A short reminder after you have read step 3A — build the binary (git clone, make) if you need to, then:
Download the small model (once):
curl -L -o models/ggml-small.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin
A minimal test command (16 kHz mono WAV; convert an MP3 with ffmpeg -i test.mp3 -ar 16000 -ac 1 test.wav):
./main -m models/ggml-small.bin -f test.wav -l fr
The transcript prints in the terminal.
Step 3 — Choose the model
| Model | Rough size | Use |
|---|---|---|
| tiny / base | Light | Tests, mostly English |
| small | ~500 MB | Everyday French, fine on CPU |
| medium | ~1.5 GB | Better French punctuation. On 8 GB of VRAM, it matters most when the file is longer than about 30 minutes |
| large | Heavy | A GPU is a good idea. On 8 GB of VRAM, skip very long files while an LLM is also running |
The first download:
whisper mon-enregistrement.mp3 --language French --model small
The weights stay on disk. Leave room on an SSD.
Step 4 — Test a local MP3
- Put
test-voix.mp3(30–60 seconds of clear speech) in a folder whose path has no accented characters (that avoids Windows errors). - Run:
whisper test-voix.mp3 --language French --model small --output_format txt
- Open
test-voix.txt. You should read a transcript in French. - If you get “ffmpeg not found”, go back to step 1.
Go further: Zoom, Teams, n8n, SRT subtitles (collapsible)
Open this only after a successful MP3 test. These ideas extend the chain. They do not replace steps 1 through 4.
Zoom / Teams: export the audio alone (MP3 or M4A), and convert it with ffmpeg before Whisper.
n8n: a watched folder, then transcription, then an Ollama summary — n8n and a local LLM.
SRT or VTT subtitles: add --output_format srt (or vtt) to the whisper command. Proofread proper names by hand.
Disk cache: clear ~/.cache/whisper (Linux and macOS) or the Windows equivalent. See SSD storage for models.
A GPU shared with Ollama: on 8 GB of VRAM, do not run Whisper large and a 7B Ollama model at the same time. Past about 30 minutes of audio, prefer medium over large as long as the quality is good enough. Schedule the jobs, or run the family assistant at a different time.
Step 5 — Summarize with Ollama or Open WebUI (optional)
A manual method (Windows, macOS, and Linux):
- Open
test-voix.txtin Notepad (Windows) or TextEdit (macOS) — or any text editor on Linux — select all, and copy. - Start Open WebUI or
ollama runwith the model you already installed. - Paste the text and ask: “Summarize this transcript as 5 bullet points in French.”
Copy and paste avoids shell tricks ($(cat …) in bash) that differ between Windows and macOS. For Zoom, n8n, or an SRT export, use the Go further block just above (after the MP3 checks out). M4A, slowness, a wrong language, or missing ffmpeg: one table, Formats and quick fixes, at the end of the guide.
Wiring this to Ollama or Open WebUI
Some Ollama versions offer speech models. The ecosystem moves, so check the current Ollama docs. For a stable stack in 2026, a dedicated Whisper plus a text LLM is still the clearest lesson: transcribe first, then summarize in Open WebUI or Docker Ollama if the family stack is already in containers.
| Step in the flow | Tool | Role |
|---|---|---|
| 1 | ffmpeg | Read MP3 and M4A |
| 2 | Whisper | Speech to a .txt file |
| 3 | Ollama / Open WebUI | Text to a structured recap |
If you record meetings with other people (coworkers, clients), keep the files on an encrypted disk and write down why you have them. See privacy for local AI before you email the .txt.
Cost and privacy
There is no subscription for open-source Whisper. The costs are electricity (a kWh budget) and, if you add it, GPU hardware (GPU, iGPU, and eGPU). Compare that with the cloud in cost over 3 years if you already pay for online tools.
Related guides (six pivots)
| Topic | Guide |
|---|---|
| Start with voice | First steps with Whisper |
| The text chain after a transcript | Open WebUI for beginners · Install Ollama |
| Privacy | Privacy when AI stays fully local |
| Audio prerequisites | The ffmpeg -version step above · a first local AI in 15 minutes |
| Automation (after the MP3 test) | The collapsible Go further block · n8n and a local LLM |
| Hardware for transcription | An SSD for models · GPU, iGPU, and eGPU |
A small model on CPU: no GPU is required for short files. A decent microphone and ffmpeg are enough.
An affiliate hardware link (Amazon affiliate, affiliation. Prices are not checked on this page, October 2026): open the link only if you are moving to medium or large, handling more than about 30 minutes per file on 8 GB of VRAM, or more than 1 hour of audio a week. An RTX 4060 (8 GB, CUDA, see the manufacturer’s spec sheet) often speeds those jobs up. See RTX 4060 cards on amazon.ca, and check today’s price before you buy.
Formats and quick fixes
| Symptom | What to try |
|---|---|
| M4A (iPhone) | ffmpeg -i voix.m4a -ar 16000 -ac 1 voix.wav, then run whisper on the WAV |
| A long MP3, or weak French punctuation | small first. Use medium if the reread is missing commas. On 8 GB of VRAM, aim for medium mainly past about 30 minutes |
| English detected by mistake | --language French |
| Slowness | tiny or small. A GPU helps for medium or large, for files longer than about 30 minutes, or for more than 1 hour a week (the hardware thresholds above) |
| Background noise | A larger model, or an ffmpeg filter first |
ffmpeg not found |
Go back to step 1 — PATH and ffmpeg -version |
| Subtitles | The Go further block (after the MP3 test) — --output_format srt or vtt |
Common questions
Which Whisper model should I use for French?
Use small or medium for a good balance on CPU. On a GPU with 8 GB of VRAM, use medium or large, especially past about 30 minutes of audio per file.
Does the transcription stay fully offline?
Yes, after the model weights are downloaded. Nothing is sent to a server if you are not using an online service.
Which audio formats are accepted?
WAV, MP3, and M4A through ffmpeg, depending on the tool. If a tool fails, convert to 16 kHz mono WAV first.
Does Whisper replace a chat LLM?
Whisper turns speech into text. To summarize that text, follow up with Ollama or Open WebUI.
Sources
Narrow it down with the hardware selector · French version of this section