Features

Subtitle Styles (Batteries Included)

The default ASS template ships with two ready-made styles, tuned for dual-language subtitles:

• Both share the classic look: white text, semi-transparent outline (3.3px) + shadow (2.5px), bottom-center aligned
• Zero-install fonts: every font is a system default — Arial ships with Windows/macOS (Linux auto-substitutes the metric-compatible Liberation Sans); Microsoft YaHei ships with Windows Chinese (macOS falls back to PingFang SC, Linux to Noto Sans CJK SC — all modern sans-serif CJK, visually near-identical)

Speaker Diarization (Who’s Talking?)

Every pipeline path can label who said what — each word gets a Speaker attribute.

Two backends, both implemented natively through the CrispASR CLI (no Python required ):

Configure via WorkflowConfig:

DiarizationBackend = "crispasr",   // "none" to disable
DiarizationMethod  = "foxnose",    // crispasr methods
DiarizationModel   = "pyannote-seg-3.0",
NumSpeakers        = 0,            // 0 = auto

Diarization runs before sentence splitting and is non-fatal: if it fails, subtitles still get generated (just without speaker labels). No drama.

Vocal Separation (BGM, Step Aside!)

Background music drowning out the speech? Separate the vocals first, then transcribe — clean input, better subtitles.

• Powered by demucs-rs (native Rust, no Python); models auto-downloaded on first run
• Mirror-safe model downloads: Demucs-rs itself only knows HuggingFace, so Centurion pre-downloads the model into its cache — and if the official source times out (we feel you, China networks ), it automatically falls back to the hf-mirror.com mirror. Zero manual steps, one working vocal track.
• Off by default — it’s slow (deep learning is patient work) and pointless for clean speech
• Turn it on only for music / MV / BGM-heavy media:

Centurion asr song.mp4 --vocal-separation
Centurion asr song.mp4 --vocal-separation --vocal-separation-model htdemucs_ft

GPU Detection & Auto-Download

Centurion detects your hardware at startup and downloads the right tool builds automatically. No more “please manually install the CUDA version” emails.

• Detects NVIDIA (via nvidia-smi / CUDA_PATH), other GPUs (AMD/Intel Vulkan), system RAM, and platform
• Tools can declare per-device variants (e.g. whisper.cpp CPU vs CUDA builds); the matching one is auto-selected & downloaded
• Override detection anytime with --device:

Centurion asr audio.mp3 --device cuda     # force CUDA builds
Centurion asr audio.mp3 --device cpu      # stay cozy on CPU

Startup banner example:

 Device: Platform: win-x64; GPU: Intel(R) Arc(TM) 130T GPU (16GB) (non-NVIDIA); RAM: 15.7 GB; Recommended: Vulkan

Metadata Registry

Tools & models live in an external JSON registry, loaded at startup — edit it without touching code.

• Default location: config/metadata.json next to the executable (seeded automatically on first run)
• Override: CENTURION_METADATA_PATH env var, or pass a path to AddCenturionCore(metadataPath)
• Tools support variants for per-device builds:

{
  "tools": {
    "whispercpp": {
      "downloadUrl": ".../whisper-bin-x64.zip",
      "variants": {
        "cuda": { "downloadUrl": ".../whisper-cublas-12.4.0-bin-x64.zip" }
      }
    }
  }
}

Non-Latin Language Support

Centurion no longer assumes your audio speaks English with spaces.

• CJK & spacing-aware text: Chinese & Japanese are joined without spaces (correct for 无空格语系), Korean keeps its word spaces — the old string.Join(" ") nightmare is dead
• Language-aware punctuation: sentence splitting recognizes 。！？，；：、… plus Devanagari ।॥ and Arabic ؟
• Per-language transcription: pass -l zh / -l ja / -l ko — Whisper & CrispASR obey; Whisper auto-detects when no language is given
• Cross-lingual alignment: the default aligner (qwen3-forced-aligner-0.6b) works for Chinese, Japanese, English & more
• LLM splitting: Chinese/Japanese use a CJK prompt branch; Korean rides the English branch (its punctuation matches anyway)
• Model-agnostic stages: diarization & vocal separation don’t care about language at all

Centurion asr 讲座.wav -l zh --transcriber qwen3-asr-1.7b
Centurion asr anime.mkv -l ja

Architecture

Operator-pipeline architecture: every stage is an independent module — swappable, extendable, reorderable. No monoliths, no tears.

 input ──► FFmpegConvert ──► AudioPreprocess ──► [ VocalSeparation] ──► Transcribe
              ──► [ Diarization] ──► SentenceSplit ──► TextCleaning ──► Alignment ──► ASS

Project Layout (clean layering, zero circular deps):

• Each step is a PipelineOperator running on a shared SubtitleWorkflowContext (immutable WorkflowConfig + mutable WorkflowState)
• Pipelines are assembled dynamically per command and executed by PipelineExecutor
• Fully async & cancellation-aware; non-fatal errors just log a warning and keep going — no half-baked bailouts

Notes & Limitations

• Active development: flags may change between versions — Centurion <command> --help is your lifesaver
• First runs download tools/models (whisper.cpp, CrispASR, Demucs, GGUF files…) — coffee recommended
• Diarization & vocal-separation models live on HuggingFace — Centurion auto-falls back to the hf-mirror.com mirror when the official source is unreachable, so the first run just works
• ⏱ Long audio + vocal separation on CPU = patience required (deep learning is worth it, we promise)