First steps with Whisper on your own PC
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.
Whisper is an open-source speech recognition model (speech-to-text: it turns speech into text). It turns an audio file into text, and the voice stays on your PC when you run it there. A chat LLM (a large language model, the kind Ollama runs) writes replies. You do not hold a conversation with Whisper. You get a transcript.
The sensible order after your first local AI in 15 minutes: once text chat works, you add voice. This tutorial covers the first transcribed file. For medium, whole folders, and more advanced ffmpeg, continue with offline French transcription with Whisper as soon as test_fr.wav shows readable French.
Who this is for
| You | Suggested path | What you need |
|---|---|---|
| Beginner, no Python, Windows or Linux | 3A whisper.cpp | Nothing installed yet. A binary download plus the model |
Comfortable with pip and a venv |
3B faster-whisper | Python 3.10+, ffmpeg |
| You would rather click than type | 3C LM Studio | LM Studio already installed, or ready to install |
Voice notes, family meetings, or an n8n chain plus a local LLM: path 3A or 3B is the straightforward one to automate.
Time (first run, not counting the model download). 3A about 25 min. 3B about 40 min (Python, a venv, and ffmpeg). 3C about 30 min if LM Studio is already in place.
Pick one path. 3A (whisper.cpp) if you want to skip Python and use a single command. 3B (faster-whisper) if you are fine with Python 3.10+. 3C (LM Studio) to drag and drop a file (comfortable on screen, and the heavier choice for an n8n batch — use 3A or 3B when you automate).
Pick a path (with Python, or without it)
| Path | Time (first run) | Choose it if… | First command |
|---|---|---|---|
| whisper.cpp (the pick when you want no Python) | ~25 min | You want an executable and one command line | Download the release plus ggml-small.bin |
| faster-whisper | ~40 min | You are fine with Python 3.10+ and pip |
pip install faster-whisper inside a venv |
| LM Studio (a graphical screen) | ~30 min | You would rather drag and drop an audio file | Install LM Studio, then the chat or audio import tab |
What you need (CPU and RAM)
| Resource | Guide |
|---|---|
| CPU | Almost any recent PC. A GPU is not required |
| RAM | 8 GB minimum. 16 GB is comfortable for medium |
| Disk | About 150 MB (small) to about 1.5 GB (medium), depending on the tool — see SSD storage for models |
| Software | ffmpeg to convert MP3 or M4A when the tool asks for it |
There is no GGUF table here (that quantized format belongs to chat models) and no VRAM table. See VRAM and which model only if you later add an LLM on the same machine.
A household workflow (text, then a summary)
- Drop your recordings in a folder named
vocaux_a_transcrire/(audio waiting to be transcribed). - Run Whisper (model small, language fr).
- Collect the transcript (a
.txtfile, or the terminal output). - Optional: send the text to Ollama for a summary — choose a first model.
- Optional: n8n watches the folder and runs transcription, then a summary.
Numbered steps
1. Prepare a test audio file
If you are only following path 3B (faster-whisper): install Python and the venv below, then continue with step 2. For 3A (whisper.cpp) or 3C (LM Studio): go straight to step 2. You do not need Python for those.
Record 30 to 60 seconds of everyday French. Name the file test_fr.wav. Common formats: WAV, MP3, M4A.
Python and a venv (step 3B only). Copy one line at a time into Command Prompt (cmd):
winget install -e --id Python.Python.3.12
python -m venv %USERPROFILE%\whisper-venv
%USERPROFILE%\whisper-venv\Scripts\activate.bat
python -m pip install --upgrade pip
PowerShell (the same thing, one line at a time; leave the $env: variables as written):
winget install -e --id Python.Python.3.12
python -m venv "$env:USERPROFILE\whisper-venv"
& "$env:USERPROFILE\whisper-venv\Scripts\Activate.ps1"
python -m pip install --upgrade pip
Linux / macOS:
python3 -m venv ~/whisper-venv && source ~/whisper-venv/bin/activate
2. Install ffmpeg (audio conversion)
-
Windows:
winget install -e –id Gyan.FFmpeg
Or the binary from ffmpeg.org.
- macOS:
brew install ffmpeg - Linux: the
ffmpegpackage from your distribution
Convert an MP3 or M4A if you need to:
ffmpeg -i note.m4a -ar 16000 -ac 1 test_fr.wav
What you want to see: at the end of step 3,
test_fr.wavproduces French text on the terminal, or in a.txtfile.
3. Transcribe (pick 3A, 3B, or 3C)
One question: are you willing to install Python 3.10+ on this PC?
- No → 3A whisper.cpp (an executable and a command line).
- Yes → 3B faster-whisper (a venv plus the
transcrire_test.pyscript below). - I want a graphical screen → 3C LM Studio (drag and drop; a weak fit for automation).
3A — whisper.cpp (~25 min, no Python)
- Download the release from whisper.cpp.
- Download
ggml-small.bin(~466 MB): the file on Hugging Face (the mirror the whisper.cpp README commonly points to). - In the folder that holds the model and the audio:
./main -m ggml-small.bin -f test_fr.wav -l fr -otxt
On Windows the executable may be named main.exe or whisper-cli.exe. Check the README in your archive. What you want to see: a file named test_fr.wav.txt, or French text printed in the terminal.
3B — faster-whisper (~40 min, Python)
With the venv active:
pip install faster-whisper
Save this script. Create transcrire_test.py in the same folder as test_fr.wav, paste the whole block below without changing the spaces, then run python transcrire_test.py:
from faster_whisper import WhisperModel
model = WhisperModel("small", device="cpu", compute_type="int8")
segments, info = model.transcribe("test_fr.wav", language="fr")
for segment in segments:
print(segment.text, end="")
print()
What you want to see: the terminal prints a French sentence or paragraph that matches your recording (punctuation can be approximate with small).
3B troubleshooting (collapsible)
If pip install faster-whisper fails, confirm the venv is active (the prompt shows (whisper-venv)). On Windows, reopen cmd or PowerShell as administrator only if winget asks for it. pip by itself does not need that.
3C — LM Studio (~30 min, a graphical screen)
3C is the comfortable on-screen path. It is handy for a one-off test, and it is the heavier path if you want a watched folder and a series of transcripts with n8n. Use 3A or 3B when you automate. LM Studio is a weak stand-in for whisper.cpp on very long files and a modest CPU. The limits, next to whisper.cpp, are also in offline French transcription with Whisper.
Follow install LM Studio, open the Chat tab, import test_fr.wav, set the language to French, and start the transcription (Transcribe / Audio).
4. Set up the incoming folder
Create vocaux_a_transcrire. Later, n8n can watch that folder, call Whisper, then send the text to Ollama — the same idea as the family assistant and school voice notes.
5. Hand the text to Ollama (optional)
- Paste the transcript into Open WebUI or
ollama run, with: “Summarize in 5 bullets in English.” - Remind the household that LLMs can invent dates. See the private assistant.
6. Check the quality
If punctuation or proper names are weak, stay on small before you step up to medium (slower on CPU). The size table is at the end of this article.
After the first successful transcript
- File the
.txtin a dated folder (transcripts/2026-10/) so your trials stay easy to find. - Compare tiny and small on the same
test_fr.wav. Note the CPU time and the mistakes on proper names. There is no need to move up to medium while small is enough for your short voice notes. - Chain it with Ollama: paste the text into Open WebUI with an instruct Q4 model (“Summarize in five bullets”).
- Automate later: the
vocaux_a_transcrire/folder plus n8n, once path 3A or 3B is stable. Leave 3C out of a batch job. - Smart home: the voice can feed Home Assistant + Ollama when the total wait (Whisper plus the LLM) is acceptable. Test first on a lamp that does not matter if it misfires.
Mistakes that show up often
| Problem | Cause | Fix |
|---|---|---|
| File not found | A path with spaces and no quotes | Put the path in quotes |
| Output in English | Language was not forced | The language="fr" setting |
| Very slow | A large model on CPU |
Switch to small |
| Codec error | An MP3 with no ffmpeg | Convert: ffmpeg -i entree.mp3 sortie.wav |
What to read next
Read these next. Once
test_fr.wavis transcribed into French, continue in this order (four guides worth opening first).
- Offline French transcription with Whisper — the medium model, several files at once, and more advanced ffmpeg.
- Your first local AI in 15 minutes — if Ollama is not installed yet and you want to summarize the transcripts.
- Open WebUI for beginners — paste the transcribed text and ask for a structured summary in English.
- n8n and a local LLM — watch
vocaux_a_transcrire/after path 3A or 3B is stable.
A microphone for your test recordings (optional)
The laptop’s built-in mic is enough for test_fr.wav. If you often record family or school voice notes, a USB microphone cuts background noise, which helps before you step up to the medium model. A sample search on Amazon (October 2026 — check the current price, the spec sheet, and whether the mic stays quiet enough for a voice note): USB microphone. Affiliate link: a commission at no extra cost to you. The rules are on the affiliate disclosure page.
Audio formats and ffmpeg
| Format | Note |
|---|---|
| WAV | Ideal for tests. The files are larger |
| MP3 | Common. ffmpeg is often required |
| M4A | iPhone messages. Convert it if you get an error |
| OGG | Open format. ffmpeg handles it well |
Example conversion:
ffmpeg -i note.m4a -ar 16000 -ac 1 test_fr.wav
Whisper is often happy with 16 kHz mono. The conversion makes CPU results more stable.
Whisper models (size versus quality)
| Model | RAM, roughly | CPU speed | French |
|---|---|---|---|
| tiny | low | fast | rough |
| base | low | fast | fine for a short clip |
| small | medium | reasonable | a good first choice |
| medium | heavier | slow | better punctuation |
| large | heavy | very slow | mainly on a GPU |
On CPU, stay on small. With an RTX you also use for Ollama, some builds accept device="cuda" for long files. Close the large chat models first (VRAM).
Hardware for transcription alone
You do not need an RTX. A 16 GB mini PC is often enough. If you stack Whisper, Ollama, and Docker, aim for 32 GB — 32/64 GB of RAM.
Safety and the GDPR (the EU General Data Protection Regulation)
After your first .txt file is transcribed locally:
- Voice notes can contain names, addresses, or a child’s voice. Treat them as sensitive documents.
- Skip copying them to a big public cloud “just to back them up” until you have read that service’s terms.
- A fully local process limits transfers to commercial processors. It still leaves the household’s duty of care in place.
- The longer frame is in privacy and fully local AI. For a cloud comparison, see the 3-year cost.
Common questions
Is Whisper a chatbot like Ollama?
Whisper turns speech into text (speech-to-text). Ollama turns text into text. The two can live on the same PC.
Which model for French on CPU?
small or medium, depending on your patience. tiny is for quick tests, and it makes more mistakes on proper names.
Do I need ffmpeg?
Usually yes, to turn MP3 or M4A into WAV or another stream the tool accepts. Install it from the official site or from your operating system’s package manager.
LM Studio or whisper.cpp for transcribing a folder automatically?
whisper.cpp or faster-whisper fit batch jobs and n8n better. LM Studio stays a comfortable path for a one-off test on screen.
Sources
Narrow it down with the hardware selector · French version of this section