Skip to content

CrispASR Pipeline

CrispASR is a self-contained external ASR engine (a C++/ggml binary from the CrispStrobe/CrispASR project). WhisperJAV shells out to it the same way Subtitle Edit does — WhisperJAV does not bundle the engine; you provide the executable.

Backends

Backend Engine Notes
parakeet (default) NVIDIA Parakeet Native word timestamps; good speed/quality
whispercpp whisper.cpp (large-v2) Native word timestamps; alias of crispasr's whisper backend
cohere Cohere Transcribe Native word timestamps; open ggml asset (no token)

All three need no forced aligner. The model GGUF auto-downloads on first use.

Prerequisites

CrispASR needs its external engine binary (not shipped with WhisperJAV):

  1. Download a release archive from CrispStrobe/CrispASR releases matching your OS and GPU:
    • Windows + NVIDIA → crispasr-windows-x86_64-cuda.zip
    • Windows, no GPU → crispasr-windows-x86_64-cpu.zip (or -vulkan)
    • macOS → crispasr-macos.tar.gz · Linux → crispasr-linux-x86_64.tar.gz
  2. Unzip it to a folder and note the path to the crispasr executable.

ffmpeg must be on PATH (already a WhisperJAV requirement). No HuggingFace token is needed for the curated backends.

How to Use

GUI (Transcription Tab)

  1. Set Mode to crispasr
  2. In the CrispASR panel: Browse to the crispasr executable, pick a Backend
  3. Click Start

GUI (Ensemble Tab)

  1. Set a pass Pipeline to CrispASR
  2. Choose the Backend in that pass's Model column; set the executable in the CrispASR panel
  3. Click Start

CLI

whisperjav video.mp4 --mode crispasr --crispasr-exe /path/to/crispasr --crispasr-backend parakeet

As an ensemble pass:

whisperjav video.mp4 --ensemble --pass1-pipeline balanced --pass2-pipeline crispasr --crispasr-exe /path/to/crispasr --crispasr-backend whispercpp

Pass advanced engine flags verbatim with --crispasr-args "...".

Where Models Are Downloaded

The engine self-manages models. On first use of a backend it downloads the GGUF (~150 MB – ~1 GB) from HuggingFace (cstr/…-GGUF) into the cache directory. WhisperJAV points CRISPASR_CACHE_DIR at ~/.cache/whisperjav/crispasr_models unless you set it yourself. Requires outbound HTTPS and curl/wget on PATH (Windows also tries WinHTTP). Subsequent runs reuse the cache.

Requirements

  • WhisperJAV installed (no extra Python package needed — the pipeline is built in)
  • ffmpeg on PATH
  • The CrispASR engine binary (user-supplied, see Prerequisites)
  • Network access on first run per backend

Strengths and Limitations

Strengths:

  • Alternative engines (Parakeet / whisper.cpp / Cohere) in one external provider
  • Self-contained — handles its own VAD; no WhisperJAV scene/segmenter/enhancer tuning needed
  • GPU variants (CUDA / Vulkan / CPU) chosen by which engine build you download

Limitations:

  • The engine binary is not bundled — you download it once yourself
  • Models pull from a mutable HuggingFace branch with no checksum (engine behavior)
  • whispercpp requests large-v2 via the engine's auto mechanism — confirm it resolves on your build (override with --crispasr-args -m <...> if needed)
  • --subs-language direct-to-english is not specially handled for CrispASR
  • First run per backend pauses for the model download

For Maintainers & Releases

  • Installer: no change required. The pipeline is pure Python inside the whisperjav package; no new pip or system dependency (FFmpeg is already required). Do not bundle the engine binary (separate project; CUDA/Vulkan/CPU variant matrix).
  • GitHub release: WhisperJAV's release assets are unchanged (the installer/wheel already contain the pipeline code). The engine binary is not a WhisperJAV asset — it lives in CrispStrobe/CrispASR releases. Models are in no release; they auto-download from HuggingFace at first use.