v1.0.3

Full Changelog: v1.0.2...v1.0.3

SubForge v1.0.2

v1.0.2 - 2026-06-02

Fixed

• Fixed Whisper.cpp word-level transcription dropping or misplacing quiet intro speech.
• Fixed smart split output covering silent/no-speech regions by running a second source-audio timing refinement after sentence reconstruction.
• Fixed subtitle timing refinement so VAD can trim both leading silence and trailing silence instead of only subtitle tails.
• Fixed RMS pause restoration cutting through Silero VAD-confirmed continuous speech in noisy driving footage.
• Fixed repeated Whisper.cpp text fragments around chunk/VAD boundaries.
• Fixed English smart-split spacing around punctuation, including cases like everyone, welcome and Torrance, California.
• Fixed dangling English fragments split after connector words such as to or this.
• Fixed packaged macOS builds so DeepFilterNet3 denoising is available in the app bundle.

Changed

• Disabled Whisper.cpp internal VAD for full-audio word timestamp runs to avoid missing quiet opening speech.
• Added live transcription and smart-split UI updates while processing.
• Improved macOS packaging for bundled torch, torchaudio, static ffmpeg/ffprobe, and denoise resources.
• Added regression tests for word-level timestamp preservation, repeated ASR text cleanup, VAD timing edge trimming, smart-split punctuation spacing, dangling-tail split avoidance, and post-split timing refinement.

Asset

• SubForge-timing-vad-fix-macos-arm64.dmg
• SHA256: 38a884b0faaaf3efdcaaf0f771d9c6e895733a830ae58c9f2e6400a36d840423

SubForge v1.0.1

Changelog

v1.0.1 - 2026-06-01

Fixed

• Fixed Whisper.cpp model selection so local Whisper.cpp transcription no longer falls through to the Jianying path.
• Fixed bundled desktop app startup and whisper.cpp binary discovery for packaged macOS builds.
• Fixed ASR task cache reuse so new transcription requests do not instantly return stale subtitle files.
• Fixed subtitle translation/optimization cache reuse so the translation page does not display old processed subtitles.
• Fixed export actions on the transcription and translation pages.
• Fixed overlapping subtitle timestamps by normalizing boundaries before export.
• Restored silence-gap handling for Whisper.cpp output while preventing adjacent subtitles from collapsing into a continuous timeline.
• Improved ASR post-processing for long subtitle spans that cover internal silence, road noise, or music.
• Added conservative sentence-final tail trimming for noisy driving footage where RMS energy stays high after speech ends.
• Fixed bilingual subtitle re-segmentation so optimized and translated subtitles keep aligned source/target lines.

Changed

• Improved local desktop packaging with bundled ffmpeg/ffprobe resources.
• Added regression tests for ASR cache behavior, Whisper.cpp routing, subtitle timestamp normalization, translation cache invalidation, optimizer cache invalidation, backend task status, and bilingual subtitle re-segmentation.

SubForge v1.0.0

SubForge v1.0.0

AI-powered video captioning tool with ASR, subtitle optimization, translation, and synthesis.

Updates (2026-05-31)

Fix: Eliminate VAD Padding-Induced Timestamp Overlaps

• Root cause: Silero VAD's speech_pad_ms (300ms default) was blindly added to both sides of every speech segment. Adjacent segments separated by less than 600ms would overlap after padding, causing subtitle timestamp overlaps in bilingual output.
• Fix: Gap-aware padding — each segment's expansion is now limited to at most half the silence gap to its neighbor. Mathematically guarantees no overlap between adjacent segments.
• Impact: Bilingual subtitles no longer have overlapping timestamps. Previously ~7.3% of subtitle segments (52/715 in a typical video) had slight overlaps.

Features

• AI Transcription: mimo-omni ASR engine for accurate speech recognition
• Intelligent Translation: LLM-powered bilingual subtitle generation (mimo v2.5-pro)
• Smart Segmentation: VAD-based silence detection with optimized subtitle splitting
• Desktop App: Native macOS application with modern web UI
• Multi-format Support: SRT, ASS, VTT subtitle export
• Audio Enhancement: DeepFilterNet3 noise reduction

Installation

macOS Desktop App

Download SubForge.dmg and drag to Applications.

CLI

pip install subforge-cli

Quick Start

# Run desktop app
open SubForge.app

# CLI usage
subforge transcribe video.mp4
subforge translate subtitles.srt --target-lang zh
subforge optimize subtitles.srt

Highlights

• Optimized VAD parameters for better silence detection
• Conservative ASR prompt to minimize hallucinations
• Position-based subtitle resegmentation
• Bilingual subtitle synthesis
• Real-time progress tracking