AutoCut: AI Audio Recording, Subtitles and Real-Time OBS Streaming

AutoCut is a local audio, subtitle, and video toolkit powered by NVIDIA CUDA, faster-whisper, Boson AI Higgs Audio v3 STT, and FFmpeg.

It provides four modular capabilities:

1. Studio Subtitle Recorder (autocut record) - Main Feature: Press record, speak naturally without streaming latency pressure, and press stop (Enter or F9) to generate high-accuracy SRT and TXT subtitles with automatic clipboard copying.
2. Real-Time Live Subtitles (autocut live): Single-pass final engine designed for OBS Studio, Twitch, YouTube, and screencasts. Outputs clean, finalized captions with zero flicker.
3. Media File Transcription (autocut transcribe): Direct transcription of existing video and audio files (MP4, MKV, MOV, WAV, MP3) into synchronized SRT and TXT subtitle files.
4. Video Silence Cutter (autocut cut): Automated dead-air and pause removal with frame-accurate FFmpeg splicing.

---

Quick Start (Windows 1-Click)

The repository root includes pre-configured batch scripts:

---

Architecture

 Microphone (sounddevice / 16 kHz)
        |
        +---> Mode 1: autocut record (Studio Recording - Main)
        |       |
        |       +---> In-memory lossless audio stream
        |       +---> RollingTranscriber: pre-emptive VAD phrase transcription on GPU
        |       +---> Instant Stop (Enter / F9): final subtitles assembled in < 0.3s
        |             |
        |             +---> captions.srt (timestamped for video editing)
        |             +---> captions.txt (clean text)
        |             +---> Windows Clipboard (ready to paste)
        |
        +---> Mode 2: autocut live (OBS Live Streaming)
        |       |
        |       +---> Adaptive noise-floor VAD
        |       +---> Single-Pass inference (finalized text only)
        |       +---> WebSocket / HTTP (:8765) -> OBS Browser Source
        |
        +---> Mode 3: autocut daemon (Headless Resident VRAM Service)
                |
                +---> Pre-warms Higgs Audio v3 STT or Whisper in GPU memory
                +---> HTTP API (:8766): /transcribe, /transcribe_chunk, /switch_model
                +---> Zero cold-start latency for all client calls

---

Main Feature: Studio Subtitle Recorder (autocut record)

Ideal for content creators, video editors (Premiere, DaVinci, CapCut), voice memos, and podcasts. Record speech without latency constraints and let the neural network transcribe the complete segment with full context.

Thanks to the RollingTranscriber pipeline, completed phrases are pre-emptively recognized in the background while you are speaking. When you stop recording, results appear almost instantaneously.

# 1-Click launch via Windows batch file:
.\run_record.bat

# Standard CLI launch (Whisper base, CUDA):
.\.venv\Scripts\autocut.exe record

# Using Higgs Audio v3 STT (LLM-decoder backend):
.\.venv\Scripts\autocut.exe record --engine higgs

# Specifying language and custom model:
.\.venv\Scripts\autocut.exe record -l en -m small.en

# Specifying custom output filename:
.\.venv\Scripts\autocut.exe record -o my_subtitles.srt

Workflow:

1. Recording starts with an active elapsed timer in the terminal.
2. Press ENTER in the console or press F9 (global hotkey across all applications) to stop recording.
3. The background worker transcribes completed phrases on-the-fly.
4. The STT engine finalizes:
   • captions.srt (timecoded subtitles)
   • captions.txt (raw transcript)
   • Automatically copied to the Windows clipboard for instant pasting.

---

Streaming Feature: Real-Time Subtitles for OBS (autocut live)

Designed for live streams and voice calls. The Single-Pass Final Engine eliminates interim draft flickering. As soon as a speaker pauses (0.4s threshold), the completed sentence is transcribed once and emitted directly as a final subtitle.

# Default fast Whisper setup (CUDA, small.en model):
.\.venv\Scripts\autocut.exe live

# Using Higgs Audio v3 STT:
.\.venv\Scripts\autocut.exe live --engine higgs

# Live translation into English:
.\.venv\Scripts\autocut.exe live --translate -l ru -m small

# Custom noise gate for loud microphones or background noise:
.\.venv\Scripts\autocut.exe live --energy-threshold 0.025

Controls During Live Streaming:

• Global Hotkey (F9): Toggles subtitle mute/pause instantly without switching windows.
• OBS Integration: Add a Browser Source pointing to http://localhost:8765/?theme=standard&size=28 with 1920x1080 resolution.

Overlay Themes:

• Standard (white text, dark plate): ?theme=standard&size=28
• Cinema Yellow: ?theme=yellow&size=28
• Outline (transparent background): ?theme=outline&size=32
• Black on White: ?theme=black&size=28
• Position customization: &pos=top or &pos=center

---

OBS Feature: Automatic Recording Synchronization (autocut obs)

Synchronizes directly with OBS Studio via the built-in OBS WebSocket protocol (OBS 28+).

How It Works:

1. AutoCut stays connected to OBS in the background.
2. Whenever you press "Start Recording" (Начать запись) in OBS:
   • AutoCut detects the recording event and starts tracking the session.
   • The STT neural network is kept pre-warmed in GPU VRAM.
3. Whenever you press "Stop Recording" (Остановить запись) in OBS:
   • AutoCut intercepts the finished video file path directly from OBS.
   • It performs offline full-context transcription using Whisper large-v3-turbo with beam_size=5 (or Higgs Audio v3 STT).
   • Generates 100% accurate synchronized .srt and .txt subtitles saved right next to the recorded video file!
   • Full transcript is automatically copied to the Windows clipboard for instant pasting.
   • Emits an audible notification chime so you know subtitles are ready even when playing in fullscreen.

# 1-Click Launch:
.\run_obs_sync.bat

# Standard CLI:
.\.venv\Scripts\autocut.exe obs

# Specify Russian / English / Auto:
.\.venv\Scripts\autocut.exe obs -l auto

# Using Higgs Audio v3 STT:
.\.venv\Scripts\autocut.exe obs --engine higgs

# Also automatically cut dead air and pauses from the finished video:
.\.venv\Scripts\autocut.exe obs --cut

Enabling WebSocket in OBS Studio:

In OBS Studio: go to Tools -> WebSocket Server Settings -> ensure "Enable WebSocket server" is checked (Server Port: 4455, default).

---

Feature 3: Resident VRAM Daemon (autocut daemon)

Starts a background local HTTP service that keeps heavy models (such as Higgs Audio v3 STT or Whisper large) resident in GPU memory to eliminate model reloading delays:

# Start daemon with Higgs Audio v3 STT:
.\.venv\Scripts\autocut.exe daemon --engine higgs --port 8766

# Start daemon with Whisper:
.\.venv\Scripts\autocut.exe daemon --engine whisper -m small.en

---

Feature 4: File Transcription (autocut transcribe)

Transcribe pre-recorded media files (MP4, MKV, MOV, WAV, MP3) into synchronized subtitles:

# Transcribe video to subtitles:
.\.venv\Scripts\autocut.exe transcribe input.mp4 -o output.srt

# Transcribe with Higgs Audio v3 STT:
.\.venv\Scripts\autocut.exe transcribe voice_memo.wav --engine higgs

---

Feature 5: Video Silence Cutter (autocut cut)

Automatically detects silence intervals and splices out dead air with sample accuracy:

# Basic silence removal:
.\.venv\Scripts\autocut.exe cut raw_video.mp4 -o edited_video.mp4

# Remove pauses longer than 0.5s with 120ms speech padding and export remapped SRT:
.\.venv\Scripts\autocut.exe cut raw.mp4 -o out.mp4 --pause-threshold 0.5 --margin 0.12 --output-srt out.srt

---

Audio Input Device Management (autocut devices)

List available audio interfaces:

.\.venv\Scripts\autocut.exe devices

Specify a device index:

.\.venv\Scripts\autocut.exe record --device-index 1
.\.venv\Scripts\autocut.exe live --device-index 1

---

Supported STT Engines

1. faster-whisper (CTranslate2):
   • Models: tiny, base, small, medium, large-v3-turbo.
   • Low VRAM footprint (500 MB - 2 GB) with CUDA float16 acceleration.
2. Higgs Audio v3 STT (bosonai/higgs-audio-v3-stt):
   • Whisper-Large-v3 acoustic encoder coupled with a Qwen3-based language model decoder.
   • High contextual understanding and punctuation accuracy.
   • Activated via the --engine higgs flag.

---

CLI Reference Summary

autocut record

Options:
  -o, --output-srt PATH       Path to export SRT subtitles [default: captions.srt]
  -e, --engine TEXT           STT engine: 'higgs' or 'whisper' [default: higgs]
  -m, --model TEXT            Whisper model name [default: large-v3-turbo]
  -d, --device TEXT           Inference device: 'cuda' or 'cpu' [default: cuda]
  --compute-type TEXT         Quantization type [default: float16]
  -l, --language TEXT         Language code ('auto', 'en', 'ru', etc.) [default: auto]
  -k, --hotkey TEXT           Hotkey to stop recording [default: f9]
  --device-index INTEGER      Microphone index from 'autocut devices'
  -w, --save-wav PATH         Optional path to save recorded audio WAV
  --clipboard / --no-clipboard Automatically copy text to clipboard [default: True]

autocut live

Options:
  -m, --model TEXT            Model name [default: small.en]
  -d, --device TEXT           Inference device [default: cuda]
  --compute-type TEXT         Quantization [default: float16]
  -l, --language TEXT         Language code [default: en]
  -T, --translate             Enable real-time spoken translation to English
  -k, --hotkey TEXT           Global hotkey to toggle pause/mute [default: f9]
  --energy-threshold FLOAT    Noise gate threshold [default: 0.015]
  -p, --port INTEGER          Server port [default: 8765]
  -o, --output-srt PATH       Path to record subtitle session [default: captions.srt]
  -t, --theme TEXT            Theme: 'standard', 'black', 'outline', 'yellow'
  -s, --size INTEGER          Font size in pixels [default: 28]
  -e, --engine TEXT           STT engine: 'whisper' or 'higgs' [default: whisper]

autocut daemon

Options:
  -p, --port INTEGER          Daemon HTTP port [default: 8766]
  -e, --engine TEXT           Pre-warmed STT engine: 'higgs' or 'whisper' [default: higgs]
  -m, --model TEXT            Model name
  -d, --device TEXT           Inference device [default: cuda]
  --compute-type TEXT         Quantization [default: float16]
  -l, --language TEXT         Default language code

autocut transcribe

Arguments:
  INPUT_FILE                  Path to audio or video file (MP4, MKV, WAV, MP3) [required]

Options:
  -o, --output-srt PATH       Path to export SRT subtitles
  -e, --engine TEXT           STT engine: 'higgs' or 'whisper' [default: higgs]
  -m, --model TEXT            Whisper model name [default: large-v3-turbo]
  -d, --device TEXT           Inference device [default: cuda]
  --compute-type TEXT         Quantization [default: float16]
  -l, --language TEXT         Language code [default: auto]
  --clipboard / --no-clipboard Copy text to Windows clipboard

autocut cut

Arguments:
  INPUT_VIDEO                 Path to input video or audio file [required]

Options:
  -o, --output PATH           Path to export trimmed video
  -m, --model TEXT            Whisper model name [default: base]
  -d, --device TEXT           Inference device [default: cuda]
  -p, --pause-threshold FLOAT Minimum silence duration in seconds [default: 0.6]
  --margin FLOAT              Audio padding margin in seconds [default: 0.15]
  -l, --language TEXT         Spoken language code
  --output-srt PATH           Export synchronized subtitles

autocut devices

Usage: autocut devices
Prints all active audio input interfaces, channels, sample rates, and default status.

---

License

MIT License. Free for personal and commercial broadcasting, recording, and editing.