FGP Whisper Daemon

Ultra-fast audio transcription with persistent Whisper model. 3-10x faster than spawning whisper per file.

Why FGP?

| Operation | FGP Daemon | whisper CLI | Speedup | |-----------|------------|-------------|---------| | Model load | 0ms (cached) | ~5-30s | ∞ | | Short audio (<30s) | 1-5s | 10-35s | 3-10x | | Long audio (5min) | 15-45s | 60-180s | 3-4x | | Batch (10 files) | 30-90s | 300-600s | 5-10x |

Persistent daemon keeps model loaded in memory (GPU or CPU).

Installation

brew install fast-gateway-protocol/tap/fgp-whisper

# Or
bash ~/.claude/skills/fgp-whisper/scripts/install.sh

Prerequisites

For local inference, install whisper.cpp or openai-whisper:

# whisper.cpp (recommended - faster, less memory)
brew install whisper-cpp

# Or Python whisper
pip install openai-whisper

# For GPU acceleration (NVIDIA)
pip install openai-whisper[cuda]

# For Apple Silicon (uses Metal)
# whisper.cpp automatically uses Metal

For API-based transcription:

export OPENAI_API_KEY="sk-..."

Usage

Basic Transcription

# Transcribe audio file
fgp whisper transcribe audio.mp3

# Transcribe video file (extracts audio)
fgp whisper transcribe video.mp4

# Output to file
fgp whisper transcribe audio.mp3 --output transcript.txt

# With timestamps
fgp whisper transcribe audio.mp3 --timestamps

# Specific language
fgp whisper transcribe audio.mp3 --language en
fgp whisper transcribe audio.mp3 --language es
fgp whisper transcribe audio.mp3 --language ja

Output Formats

# Plain text (default)
fgp whisper transcribe audio.mp3 --format txt

# JSON with segments
fgp whisper transcribe audio.mp3 --format json

# SRT subtitles
fgp whisper transcribe audio.mp3 --format srt

# VTT subtitles
fgp whisper transcribe audio.mp3 --format vtt

# TSV (tab-separated)
fgp whisper transcribe audio.mp3 --format tsv

# All formats
fgp whisper transcribe audio.mp3 --format all

Models

# List available models
fgp whisper models

# Use specific model
fgp whisper transcribe audio.mp3 --model tiny      # Fastest, least accurate
fgp whisper transcribe audio.mp3 --model base      # Fast, good accuracy
fgp whisper transcribe audio.mp3 --model small     # Balanced (default)
fgp whisper transcribe audio.mp3 --model medium    # Better accuracy
fgp whisper transcribe audio.mp3 --model large     # Best accuracy, slowest
fgp whisper transcribe audio.mp3 --model large-v3  # Latest large model

# English-only models (faster)
fgp whisper transcribe audio.mp3 --model tiny.en
fgp whisper transcribe audio.mp3 --model base.en
fgp whisper transcribe audio.mp3 --model small.en
fgp whisper transcribe audio.mp3 --model medium.en

# Download model
fgp whisper download large-v3

# Check loaded model
fgp whisper status

Translation

# Translate to English
fgp whisper translate audio.mp3

# Translate from specific language
fgp whisper translate audio.mp3 --from es

# Translate with subtitles
fgp whisper translate video.mp4 --format srt

Language Detection

# Detect language
fgp whisper detect audio.mp3

# Get confidence scores
fgp whisper detect audio.mp3 --verbose

Batch Processing

# Transcribe multiple files
fgp whisper batch *.mp3

# Transcribe folder
fgp whisper batch ./audio --output ./transcripts

# With specific format
fgp whisper batch ./audio --format srt --output ./subtitles

# Parallel processing
fgp whisper batch ./audio --parallel 4

Real-time / Streaming

# Transcribe from microphone
fgp whisper stream

# Stream to file
fgp whisper stream --output live.txt

# Stream with translation
fgp whisper stream --translate

# Set audio device
fgp whisper stream --device 1

Timestamps & Segments

# Word-level timestamps
fgp whisper transcribe audio.mp3 --word-timestamps

# Segment by sentence
fgp whisper transcribe audio.mp3 --segment sentence

# Custom segment length
fgp whisper transcribe audio.mp3 --segment-length 30

# Get specific time range
fgp whisper transcribe audio.mp3 --start 00:01:30 --end 00:05:00

Speaker Diarization

# Identify speakers
fgp whisper transcribe audio.mp3 --diarize

# Set number of speakers
fgp whisper transcribe audio.mp3 --diarize --speakers 2

# With speaker labels
fgp whisper transcribe audio.mp3 --diarize --speaker-labels "Alice,Bob"

Audio Processing

# Normalize audio before transcription
fgp whisper transcribe audio.mp3 --normalize

# Reduce noise
fgp whisper transcribe audio.mp3 --denoise

# Specific sample rate
fgp whisper transcribe audio.mp3 --sample-rate 16000

# Mono conversion
fgp whisper transcribe audio.mp3 --mono

OpenAI API

# Use OpenAI API instead of local model
fgp whisper transcribe audio.mp3 --api openai

# With specific model
fgp whisper transcribe audio.mp3 --api openai --model whisper-1

# Set API key
export OPENAI_API_KEY="sk-..."

Prompting

# Initial prompt for context
fgp whisper transcribe audio.mp3 --prompt "This is a podcast about technology."

# Technical vocabulary
fgp whisper transcribe audio.mp3 --prompt "Kubernetes, Docker, PostgreSQL, GraphQL"

# Prompt from file
fgp whisper transcribe audio.mp3 --prompt-file context.txt

Performance Tuning

# CPU threads
fgp whisper transcribe audio.mp3 --threads 8

# GPU acceleration
fgp whisper transcribe audio.mp3 --device cuda

# Apple Metal (default on macOS)
fgp whisper transcribe audio.mp3 --device metal

# Beam size (accuracy vs speed)
fgp whisper transcribe audio.mp3 --beam-size 5

# Temperature (randomness)
fgp whisper transcribe audio.mp3 --temperature 0

Available Commands

| Command | Description | Example | |---------|-------------|---------| | transcribe | Transcribe audio | fgp whisper transcribe audio.mp3 | | translate | Translate to English | fgp whisper translate audio.mp3 | | detect | Detect language | fgp whisper detect audio.mp3 | | batch | Batch transcribe | fgp whisper batch *.mp3 | | stream | Real-time from mic | fgp whisper stream | | models | List models | fgp whisper models | | download | Download model | fgp whisper download large-v3 | | status | Daemon status | fgp whisper status |

Supported Formats

Audio:

• MP3, WAV, FLAC, OGG, M4A, AAC, WMA, AIFF

Video (audio extracted):

• MP4, MKV, AVI, MOV, WebM, FLV

Output:

• TXT, JSON, SRT, VTT, TSV

Supported Languages

Whisper supports 99 languages including:

| Code | Language | Code | Language | |------|----------|------|----------| | en | English | es | Spanish | | fr | French | de | German | | it | Italian | pt | Portuguese | | ru | Russian | ja | Japanese | | ko | Korean | zh | Chinese | | ar | Arabic | hi | Hindi |

Full list: fgp whisper languages

Example Workflows

Podcast transcription

# Transcribe podcast with speaker diarization
fgp whisper transcribe podcast.mp3 \
  --model medium \
  --diarize \
  --format json \
  --output podcast_transcript.json

Video subtitles

# Generate SRT subtitles for video
fgp whisper transcribe video.mp4 \
  --format srt \
  --output video.srt

# For non-English video, translate subtitles
fgp whisper translate video.mp4 \
  --format srt \
  --output video_en.srt

Meeting notes

# Transcribe meeting with timestamps
fgp whisper transcribe meeting.m4a \
  --model small \
  --timestamps \
  --diarize \
  --output meeting_notes.txt

Batch process interviews

# Transcribe all interviews
fgp whisper batch ./interviews \
  --model medium \
  --format json \
  --diarize \
  --parallel 2 \
  --output ./transcripts

Live captioning

# Real-time captions from microphone
fgp whisper stream --model base.en

Troubleshooting

Model not found

Error: Model 'large-v3' not found

Download model: fgp whisper download large-v3

Out of memory

Error: CUDA out of memory

Use smaller model: --model small or use CPU: --device cpu

Audio format error

Error: Unsupported audio format

Install ffmpeg: brew install ffmpeg

Slow transcription

Tip: Transcription is slow

• Use GPU acceleration: --device cuda or --device metal
• Use smaller model: --model base or --model tiny
• Use English-only model: --model small.en

Inaccurate results

Tip: Improve accuracy

• Use larger model: --model large-v3
• Add context prompt: --prompt "Technical terms..."
• Check audio quality

Architecture

• whisper.cpp or openai-whisper backend
• Model caching in memory (GPU or CPU)
• UNIX socket at ~/.fgp/services/whisper/daemon.sock
• Metal acceleration on Apple Silicon
• CUDA acceleration on NVIDIA GPUs
• Streaming support for real-time transcription