Stt Transcription

Speech-to-Text Transcription

Comprehensive speech-to-text capabilities using multiple STT engines. Record audio, transcribe files, real-time processing, speaker identification, and multi-language support.

Quick Start

When asked to transcribe audio:

1. Choose engine: Whisper (local/free), Google, Azure, or AssemblyAI
2. Record or load: Capture audio or use existing file
3. Transcribe: Convert speech to text
4. Format: Output as plain text, SRT, VTT, or JSON
5. Enhance: Add timestamps, speaker labels, punctuation

Prerequisites

System Requirements

• Python 3.8+
• Microphone (for recording)
• Audio file support: WAV, MP3, M4A, FLAC, OGG

Install Dependencies

Core (required):

pip install sounddevice soundfile numpy --break-system-packages

Whisper (OpenAI - local, free):

pip install openai-whisper --break-system-packages
# For faster processing with GPU:
pip install openai-whisper torch --break-system-packages

Google Speech (requires API key):

pip install google-cloud-speech --break-system-packages

Azure Speech (requires API key):

pip install azure-cognitiveservices-speech --break-system-packages

AssemblyAI (requires API key):

pip install assemblyai --break-system-packages

Optional enhancements:

pip install pydub webrtcvad --break-system-packages  # Audio processing
pip install pyaudio --break-system-packages  # Alternative audio backend

See reference/setup-guide.md for detailed installation.

STT Engine Comparison

Whisper (Recommended for most users)

• ✅ Free and local - No API costs, runs offline
• ✅ High quality - State-of-the-art accuracy
• ✅ Multilingual - 99+ languages
• ⚠️ Speed - Slower than cloud services (depends on hardware)
• ⚠️ Resources - Needs decent CPU/GPU

Google Cloud Speech

• ✅ Fast - Real-time capable
• ✅ Accurate - Excellent for English
• ✅ Features - Automatic punctuation, speaker diarization
• ⚠️ Cost - $0.006 per 15 seconds (~$1.44/hour)
• ⚠️ Privacy - Audio sent to Google

Azure Speech

• ✅ Enterprise - Microsoft integration
• ✅ Translation - Real-time translation
• ✅ Custom - Train custom models
• ⚠️ Cost - $1 per audio hour
• ⚠️ Setup - More complex configuration

AssemblyAI

• ✅ Features - Speaker diarization, sentiment analysis
• ✅ Quality - Very accurate
• ✅ Developer-friendly - Simple API
• ⚠️ Cost - $0.00025 per second (~$0.90/hour)

Core Operations

Record Audio

Simple recording:

# Record 30 seconds
python scripts/record_audio.py --duration 30 --output recording.wav

# Record until stopped (Ctrl+C)
python scripts/record_audio.py --output recording.wav

# Record with voice activity detection
python scripts/record_audio.py --vad --output recording.wav

Advanced recording:

# Choose microphone
python scripts/list_devices.py  # List available mics
python scripts/record_audio.py --device 1 --output recording.wav

# Specify quality
python scripts/record_audio.py \
  --sample-rate 48000 \
  --channels 2 \
  --output recording.wav

Transcribe Files

Using Whisper (local, free):

# Basic transcription
python scripts/transcribe_whisper.py --file recording.wav

# Choose model size (tiny, base, small, medium, large)
python scripts/transcribe_whisper.py \
  --file recording.wav \
  --model medium

# With timestamps
python scripts/transcribe_whisper.py \
  --file recording.wav \
  --timestamps \
  --output transcript.json

# Multiple languages
python scripts/transcribe_whisper.py \
  --file recording.wav \
  --language es  # Spanish

Using Google Cloud:

# Export API key
export GOOGLE_APPLICATION_CREDENTIALS="path/to/credentials.json"

# Transcribe
python scripts/transcribe_google.py \
  --file recording.wav \
  --language en-US

# With speaker diarization
python scripts/transcribe_google.py \
  --file recording.wav \
  --diarization \
  --speakers 2

Using Azure:

# Set credentials
export AZURE_SPEECH_KEY="your-key"
export AZURE_SPEECH_REGION="westus"

# Transcribe
python scripts/transcribe_azure.py --file recording.wav

# Real-time
python scripts/transcribe_azure_realtime.py --microphone

Using AssemblyAI:

# Set API key
export ASSEMBLYAI_API_KEY="your-key"

# Transcribe with features
python scripts/transcribe_assemblyai.py \
  --file recording.wav \
  --diarization \
  --sentiment \
  --topics

Real-Time Transcription

Stream from microphone:

# Whisper streaming (chunked)
python scripts/stream_whisper.py

# Google streaming
python scripts/stream_google.py

# Azure continuous recognition
python scripts/stream_azure.py

Format Output

Plain text:

python scripts/transcribe_whisper.py --file audio.wav --output transcript.txt

JSON with metadata:

python scripts/transcribe_whisper.py \
  --file audio.wav \
  --format json \
  --output transcript.json

# Output includes:
# - Text segments
# - Timestamps
# - Confidence scores
# - Language detection

SRT subtitles:

python scripts/transcribe_whisper.py \
  --file video.mp4 \
  --format srt \
  --output subtitles.srt

VTT subtitles:

python scripts/transcribe_whisper.py \
  --file video.mp4 \
  --format vtt \
  --output subtitles.vtt

Common Workflows

Workflow 1: Meeting Transcription

Scenario: Record and transcribe meeting with speaker labels

# 1. Record meeting
python scripts/record_audio.py \
  --output meeting.wav \
  --vad  # Stop on silence

# 2. Transcribe with speaker diarization
python scripts/transcribe_google.py \
  --file meeting.wav \
  --diarization \
  --speakers 4 \
  --output meeting.json

# 3. Format for readability
python scripts/format_transcript.py \
  --input meeting.json \
  --format markdown \
  --output meeting.md

# Result: Formatted transcript with speaker labels and timestamps

Workflow 2: Voice Notes to Markdown

Scenario: Quick voice note → markdown document

# Record voice note
python scripts/quick_note.py

# (Records audio, transcribes with Whisper, saves as markdown)
# Output: voice-note-2025-01-20-14-30.md

Workflow 3: Batch Transcription

Scenario: Transcribe multiple audio files

# Batch process folder
python scripts/batch_transcribe.py \
  --input ./recordings/ \
  --output ./transcripts/ \
  --engine whisper \
  --model base

# Progress shown for each file

Workflow 4: Video Subtitles

Scenario: Generate subtitles for video

# Extract audio from video
python scripts/extract_audio.py --video lecture.mp4 --output audio.wav

# Generate subtitles
python scripts/transcribe_whisper.py \
  --file audio.wav \
  --format srt \
  --output lecture.srt

# Embed in video (requires ffmpeg)
python scripts/embed_subtitles.py \
  --video lecture.mp4 \
  --subtitles lecture.srt \
  --output lecture-subbed.mp4

Workflow 5: Multi-Language Support

Scenario: Transcribe and translate

# Transcribe Spanish audio
python scripts/transcribe_whisper.py \
  --file spanish-audio.wav \
  --language es \
  --output transcript-es.txt

# Translate to English
python scripts/transcribe_whisper.py \
  --file spanish-audio.wav \
  --task translate \
  --output transcript-en.txt

Whisper Model Sizes

Recommendation:

• Casual use: tiny or base (fast, good enough)
• Quality needed: small or medium (balanced)
• Professional: large (best accuracy, slower)
• GPU available: Use medium or large
• CPU only: Use tiny or base

Language Support

Whisper supports 99+ languages:

# Common languages
en  # English
es  # Spanish
fr  # French
de  # German
it  # Italian
pt  # Portuguese
nl  # Dutch
pl  # Polish
ru  # Russian
ja  # Japanese
ko  # Korean
zh  # Chinese
ar  # Arabic
hi  # Hindi

Full list: reference/language-codes.md

Speaker Diarization

Identify who said what:

# Google (best diarization)
python scripts/transcribe_google.py \
  --file meeting.wav \
  --diarization \
  --speakers 3  # Hint: 3 speakers expected

# AssemblyAI
python scripts/transcribe_assemblyai.py \
  --file meeting.wav \
  --diarization

# Output format:
# Speaker 1: Hello everyone, let's begin
# Speaker 2: Thanks for joining
# Speaker 1: Today's agenda includes...

Post-process with names:

python scripts/label_speakers.py \
  --transcript meeting.json \
  --labels "Alice,Bob,Charlie" \
  --output meeting-labeled.txt

Audio Processing

Enhance audio quality:

# Reduce noise
python scripts/denoise_audio.py \
  --input noisy.wav \
  --output clean.wav

# Normalize volume
python scripts/normalize_audio.py \
  --input quiet.wav \
  --output normalized.wav

# Convert format
python scripts/convert_audio.py \
  --input audio.m4a \
  --output audio.wav

Timestamps and Segments

Transcript with timestamps:

{
  "segments": [
    {
      "start": 0.0,
      "end": 3.5,
      "text": "Welcome to today's meeting.",
      "confidence": 0.95
    },
    {
      "start": 3.5,
      "end": 7.2,
      "text": "Let's review the quarterly results.",
      "confidence": 0.92
    }
  ]
}

Search by timestamp:

# Find text at specific time
python scripts/find_at_time.py \
  --transcript meeting.json \
  --time "5:30"  # 5 minutes 30 seconds

# Extract time range
python scripts/extract_range.py \
  --transcript meeting.json \
  --start "2:00" \
  --end "5:00" \
  --output excerpt.txt

API Cost Comparison

Per hour of audio:

• Whisper: Free (local processing)
• Google: ~$1.44 (60 min × $0.024/min)
• Azure: ~$1.00 (standard pricing)
• AssemblyAI: ~$0.90 (3600 sec × $0.00025/sec)

Free tiers:

• Google: $300 credit (first 90 days)
• Azure: 5 hours/month free
• AssemblyAI: 3 hours free on signup

Scripts Reference

Recording:

• record_audio.py - Record from microphone
• list_devices.py - List audio devices
• test_microphone.py - Test mic input

Transcription:

• transcribe_whisper.py - Whisper transcription
• transcribe_google.py - Google Cloud STT
• transcribe_azure.py - Azure Speech
• transcribe_assemblyai.py - AssemblyAI

Real-time:

• stream_whisper.py - Whisper streaming
• stream_google.py - Google streaming
• stream_azure.py - Azure continuous

Processing:

• batch_transcribe.py - Batch processing
• format_transcript.py - Format output
• extract_audio.py - Extract from video
• denoise_audio.py - Noise reduction

Utilities:

• quick_note.py - Record + transcribe
• label_speakers.py - Add speaker names
• find_at_time.py - Search by timestamp
• convert_audio.py - Format conversion

Best Practices

1. Start with Whisper - Free, offline, good quality
2. Test different models - Balance speed vs accuracy
3. Use VAD - Voice Activity Detection for cleaner recording
4. Enhance audio first - Denoise for better results
5. Appropriate model size - Don't use large models for quick notes
6. Speaker diarization - Essential for meetings
7. Save raw audio - Keep original for re-processing
8. Add context - Language hints improve accuracy

Troubleshooting

"No module named 'whisper'"

pip install openai-whisper --break-system-packages

"Microphone not working"

# List devices
python scripts/list_devices.py

# Test specific device
python scripts/test_microphone.py --device 1

"Out of memory" (Whisper)

# Use smaller model
python scripts/transcribe_whisper.py --file audio.wav --model tiny

# Or process in chunks
python scripts/transcribe_chunked.py --file large-audio.wav

"Poor transcription quality"

• Use larger Whisper model (medium/large)
• Enhance audio first (denoise, normalize)
• Specify correct language
• Check microphone quality

"API authentication failed"

# Google
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/credentials.json"

# Azure
export AZURE_SPEECH_KEY="your-key"
export AZURE_SPEECH_REGION="region"

# AssemblyAI
export ASSEMBLYAI_API_KEY="your-key"

Integration Examples

See examples/ for complete workflows:

• examples/meeting-minutes.md - Meeting transcription
• examples/podcast-notes.md - Podcast processing
• examples/lecture-subtitles.md - Video subtitles
• examples/voice-journal.md - Voice note system

Reference Documentation

• reference/setup-guide.md - Detailed setup
• reference/engine-comparison.md - STT engine details
• reference/language-codes.md - Supported languages
• reference/api-keys.md - Getting API credentials
• reference/audio-formats.md - Format specifications