Whisper Project v1.7.0

Whisper Project v1.7.0

A settings-safety and stability release on top of v1.6.0.

Highlights

• New "Enable VAD (skip silent segments)" checkbox (Advanced >
  transcription settings). Voice Activity Detection was always on with no
  way to turn it off from the desktop app. The three VAD tuning sliders
  (min silence, threshold, speech pad) now grey out while it is unchecked.
• Transcript viewer: "Edit timestamp..." — right-click a segment to
  hand-edit its start/end time. Only that one segment changes. A resulting
  overlap with the previous segment, or a sub-1-second duration, gets a
  light-orange row highlight as a warning; it never blocks saving.

Fixed

• Editing "Hotwords" in Advanced settings could be silently undone.
  Saving Advanced settings correctly wrote the new hotwords to disk, but
  the Transcribe tab kept its own frozen copy from app launch. Touching
  the language dropdown or a checkbox on that tab afterward silently
  overwrote the fresh edit. Fixed.
• config.json now refuses a drastic silent shrink and keeps a
  backup. A real incident this cycle reduced a user's config from ~90
  keys to 3 with no clear trigger found. Saving now refuses to write a
  config with under 40% of the key count currently on disk, and keeps a
  rotating config.json.bak of the last good state on every normal save.
• A rare native crash during engine/hardware detection is now
  hardened against. A background thread probing for an optional heavy
  package (Google Cloud Speech, whisper.cpp, a torch-backed engine,
  pyannote.audio, sentence-transformers, llama-cpp-python, sherpa-onnx,
  stable-ts, or the hardware tier probe) could occasionally crash the
  whole app if Python's garbage collector ran at the wrong moment during
  the import. Every such probe now disables garbage collection for that
  one operation, under one shared lock so two probes on different
  threads cannot undo each other's protection.
• Semantic search could silently mis-score a stale index row after
  switching embedding models. It now skips a dimension-mismatched row
  instead of scoring a meaningless partial comparison.

Security

• "Convert transcript" ELAN (.eaf) import now rejects a DOCTYPE. A
  crafted .eaf file with a DOCTYPE/entity block could previously hang
  or balloon memory ("billion laughs") when imported. Such files are now
  refused outright with a clear error.

Other

• The installer and Portable build no longer bundle the
  google-cloud-speech / google-cloud-storage client libraries. Google
  Cloud STT is an opt-in engine for users with their own service-account
  key; those libraries now install on demand the first time it is picked,
  the same as every other optional backend. This shrinks the normal
  install.

Builds

• Setup-Standard (Windows) — the recommended installer (embeddable
  Python; choose where models are stored on first run).
• Portable (Windows) — a ZIP of the same tree; extract and run
  Run Whisper Project.bat, no install.
• macOS — WhisperProject-v1.5.0-macOS-x64.dmg (Intel/x64-only) is
  attached to this release too, carried forward from v1.5.0 unchanged;
  no new macOS build was done this round, so it does not include any
  v1.6.0/v1.7.0 fix.

Notes

• First launch asks where to keep the speech models (large files); the
  default is a writable per-user folder.
• Windows SmartScreen may warn on an unsigned installer — choose More
  info → Run anyway.

Whisper Project v1.6.0

Whisper Project v1.6.0

A Live-transcription + denoise release on top of v1.5.0.

Highlights

• New Live tab. Transcribe a microphone or the system audio as it
  happens. Text appears while you speak instead of after a file finishes.
  Chunks are cut at natural pauses so words are not split in half, silence
  is never sent to the model, and the tab says so when the machine cannot
  keep up rather than skipping audio silently. System-audio capture is
  Windows-only. See docs/LIVE.md.
• Adaptive audio denoise before transcription (Advanced > AI Layer,
  off by default). Cuts hallucinated lines and misheard words on noisy
  recordings. It measures each file first, leaves already-clean audio
  untouched, then checks its own output and falls back to the original if
  the filter removed speech instead of noise. Bundled ffmpeg only — no
  extra download. See docs/DENOISE.md.
• ASS / SSA subtitle support. A new ass output format, and .ass /
  .ssa files can now be converted from as well. ASS is what video editors
  and karaoke tools expect, and it is the first format that carries our
  per-word timings as real karaoke highlighting.
• platform\windows\update.bat — a one-command updater for a source
  (git clone) install on Windows.
• The README is now available in 7 more languages — Chinese, Japanese,
  Korean, German, Spanish, French, Portuguese — reachable from a flag
  switcher at the top of each one.

Fixed

• The nvidia_asr (Parakeet) backend could fail to import with a
  tokenizers>=X,<=Y is required error on a build done the wrong day.
  tokenizers and transformers are now pinned to a verified matching
  pair, and the backend's error message and status probe both report the
  real cause.
• The default engine is always offline faster-whisper again. It no
  longer flips to Google Cloud STT because a key file exists next to the
  app — that bundled key was revoked; see SECURITY.md.
• Resuming a cancelled job now re-checks the pre-processing settings.
  Changing vocal separation or denoise between cancel and resume used to
  splice differently-conditioned halves into one transcript; the partial
  is now invalidated and re-run instead.

Builds

• Setup-Standard (Windows) — the recommended installer (embeddable
  Python; choose where models are stored on first run).
• Portable (Windows) — a ZIP of the same tree; extract and run
  Run Whisper Project.bat, no install.

macOS: no new build this round. The v1.5.0 release still has
WhisperProject-v1.5.0-macOS-x64.dmg (Intel/x64-only) for Mac users
until a future session rebuilds it.

Notes

• First launch asks where to keep the speech models (large files); the
  default is a writable per-user folder.
• Windows SmartScreen may warn on an unsigned installer — choose More
  info → Run anyway.
• The v1.5.0 release stays published (not deleted) alongside this one.

Whisper Project v1.5.0

Whisper Project v1.5.0

An SMTV release on top of v1.4.0.

Highlights

• SMTV docx header now shows the detected language. Row 2 / column 3 used
  to always read the literal "Foreign Language"; it now shows the language
  faster-whisper detected (e.g. "Korean"), matching the title row and the
  "[... starts]" cue. With no detected language the header keeps its
  original generic text.
• SMTV added to File → Convert transcript. The format picker now offers
  smtv_docx alongside the existing text targets, for turning any already
  -produced transcript into the team's SMTV template.
• Project renamed to whisper_app (GitHub repo + local folder).

Builds

• Setup-Standard (Windows) — the recommended installer (embeddable
  Python; choose where models are stored on first run).
• Portable (Windows) — a ZIP of the same tree; extract and run
  Run Whisper Project.bat, no install.

This release is Windows-only; macOS builds resume in a future release.

Notes

• First launch asks where to keep the speech models (large files); the
  default is a writable per-user folder.
• Windows SmartScreen may warn on an unsigned installer — choose More info
  → Run anyway.

basic v0.1.0 — first release of the simplified edition

Whisper Project — basic v0.1.0

First release of the basic edition — a radically simplified
fork of the full-fat Whisper Project. One screen, great defaults,
no airplane-dashboard of options.

This release ships as a single Windows installer with an embedded
Python interpreter (no system Python required, no Portable EXE).

What it does

• Transcribe — drag-and-drop or Browse a media file, click
  Transcribe. Runs faster-whisper
  large-v3 locally. Writes .srt, .json, and .txt next to
  the source.
• Language picker — Auto-detect (default), English, Chinese,
  or Vietnamese. Anything else still works via auto-detect.
• Download Videos (optional) — paste one URL per line; the
  bundled yt-dlp.exe (or the SMTV scraper for Supreme Master TV
  URLs) downloads to your chosen folder. Optional time-range
  slice; optional auto-transcribe after download.
• Cancel any running job (Esc or right-click).
• Hardware autodetect — CUDA when available, otherwise CPU.

What it doesn't do (by design)

• No tabs, no Advanced dialog, no theme switcher.
• No transcript viewer / search / history / watched folder / tray.
• No diarisation, chapters, voiceprint, LLM, Demucs, alignment.
• No alternative backends (whisper.cpp, Parakeet) — only
  faster-whisper.
• No DOCX / PDF / LRC / VTT / TSV writers.
• No sponsorblock / subtitle download / burn-in.

For any of those, use the full-fat repo's
chore/cleanup-hardening
branch.

Self-diagnostics

If anything misbehaves:

• Help → Diagnose — re-runs the startup checks; copy the
  report to clipboard.
• Help → Show recent log — last 200 lines of the app log.
• Help → Open log folder — opens the OS file manager on
  %LOCALAPPDATA%\WhisperProjectBasic\Logs\.
• Unhandled exceptions surface a crash dialog with full
  traceback + Copy + Open log folder buttons — never silent.

Install

1. Download WhisperProjectBasic-v0.1.0-Setup.exe from the
   Assets below.
2. Double-click; click Yes on the UAC prompt.
3. Confirm install location (default C:\Program Files\WhisperProjectBasic\).
4. Optionally tick "Create a desktop icon".
5. Launch from Start Menu under Whisper Project (basic).

The first launch shows a hub-folder picker (where Whisper models
will live) — accept the default <install>\hub or pick a roomy
drive. The first Transcribe click triggers a one-off ~3 GB
model download from the CDN (MD5-verified). Subsequent launches
load the model from disk in ~15 seconds.

Quality bar

Metric	Result
pyright app/ core/	0 errors, 0 warnings, 0 informations
Unit + integration suite	998 tests passing in ~10 s
Real-file end-to-end (90 s English clip, CPU)	106 s — done event, three output files
Adversarial runtime probe	26 scenarios, 23 PASS + 3 WARN — both real WARNs fixed
Hostile code audit	39 findings catalogued, 27 fixed (7 P0 + 20 P1)

The full audit + probe reports are committed at the repo root
(DEBUG_AUDIT_BASIC.md, DEBUG_RUNTIME_BASIC.md).

Asset

File	Size
WhisperProjectBasic-v0.1.0-Setup.exe	~161 MB

Uninstall via Settings → Apps → Whisper Project (basic). If
your hub folder lives outside the install dir, the uninstaller
asks before deleting your (potentially multi-GB) model files.