Whisper: Local Speech-to-Text
Without the Cloud

# Requires Python 3.8+ and ffmpeg

# macOS
brew install ffmpeg
pip install openai-whisper

# Linux (Ubuntu/Debian)
sudo apt install ffmpeg
pip install openai-whisper

# Windows (with chocolatey)
choco install ffmpeg
pip install openai-whisper

# Transcribe a file (model downloads automatically on first run)
whisper meeting.mp3 --model medium

# Specify output formats
whisper interview.m4a --model medium --output_format txt

# Generate subtitles (SRT format)
whisper lecture.mp4 --model medium --output_format srt

# Non-English audio — specify language for better accuracy
whisper paris-meeting.mp3 --model medium --language French

# Translate to English while transcribing
whisper german-call.mp3 --model large-v3 --task translate

# Best accuracy on capable hardware
whisper important-meeting.mp3 --model large-v3-turbo

pip install faster-whisper

from faster_whisper import WhisperModel

# Use "cuda" if you have an NVIDIA GPU, "cpu" otherwise
# "auto" selects automatically
model = WhisperModel(
    "large-v3-turbo",
    device="auto",
    compute_type="int8"       # int8 = fastest, lowest RAM
)

segments, info = model.transcribe(
    "meeting.mp3",
    beam_size=5
)

print(f"Detected language: {info.language} (probability: {info.language_probability:.2f})")

for segment in segments:
    print(f"[{segment.start:.1f}s → {segment.end:.1f}s] {segment.text}")

# Build from source
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make

# Download a model (GGML format)
bash models/download-ggml-model.sh medium

# Convert audio to 16kHz WAV (required format)
ffmpeg -i meeting.mp3 -ar 16000 -ac 1 -c:a pcm_s16le meeting.wav

# Transcribe (pure CPU, no Python)
./main -m models/ggml-medium.bin -f meeting.wav

# Apple Silicon: enable Core ML acceleration
make -j WHISPER_COREML=1
./main -m models/ggml-large-v3-turbo.bin -f meeting.wav

from faster_whisper import WhisperModel
import ollama
import sys

# ── Step 1: Transcribe locally with Whisper ──────────────────
print("🎙️  Transcribing... (audio never leaves this machine)")

whisper_model = WhisperModel("large-v3-turbo", device="auto", compute_type="int8")
audio_file = sys.argv[1]  # e.g. python pipeline.py meeting.mp3

segments, info = whisper_model.transcribe(audio_file, beam_size=5)

transcript = ""
for segment in segments:
    transcript += f"[{segment.start:.0f}s] {segment.text.strip()}\n"

print(f"\n✅ Transcribed {info.language} audio ({len(transcript)} chars)\n")

# ── Step 2: Summarise locally with Ollama ───────────────────
print("🧠  Summarising with local LLM...")

response = ollama.chat(
    model="nous-hermes-2",   # or openchat, llama3, etc.
    messages=[{
        'role': 'user',
        'content': f"""Summarise this meeting transcript. 
Extract: (1) key decisions, (2) action items with owners, 
(3) open questions.

TRANSCRIPT:
{transcript}"""
    }]
)

print("\n📋 Meeting Summary:")
print(response['message']['content'])

# Both steps ran entirely locally.
# No audio and no text left this machine.

# Transcribe every audio/video file in a directory
for f in ~/recordings/*.{mp3,mp4,m4a,wav,ogg}; do
    echo "Processing: $f"
    whisper "$f" --model medium --output_dir ~/transcripts/
done

pip install faster-whisper pyannote.audio

# pyannote requires accepting terms at hf.co/pyannote/speaker-diarization-3.1
# then use your HuggingFace token:
# export HF_TOKEN="hf_..."

from faster_whisper import WhisperModel
from pyannote.audio import Pipeline
import os

audio_file = "meeting.wav"  # must be 16kHz mono WAV

# ── Speaker diarisation ────────────────────────────────────
diarize = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    use_auth_token=os.getenv("HF_TOKEN")
)
diarization = diarize(audio_file)

# ── Transcription ──────────────────────────────────────────
whisper = WhisperModel("medium", device="auto", compute_type="int8")
segments, _ = whisper.transcribe(audio_file, beam_size=5)
segments = list(segments)

# ── Align segments with speakers ──────────────────────────
for segment in segments:
    mid = (segment.start + segment.end) / 2
    speaker = "Unknown"
    for turn, _, spk in diarization.itertracks(yield_label=True):
        if turn.start <= mid <= turn.end:
            speaker = spk
            break
    print(f"[{speaker}] {segment.text.strip()}")