Skip to content

Transcription Engines

Subcast ships two local transcription engines. The default auto mode dispatches by audio language.

Engine comparison

SenseVoiceWhisper
Languageszh / en / ja / ko / yue99 languages
Speed~15x faster on CPUSlower (worse with bigger tiers)
Chinese accuracyExcellentMediocre at lower tiers
Cue granularitySentence-level (VAD segments)Word-level timestamps
Model size250 MB (bundled)77 MB – 1.6 GB (on demand)

Auto dispatch

The default mode. At the start of each task, the opening seconds are sampled and the language voted on:

  • Chinese / Japanese / Korean → SenseVoice (language locked for the task — fast and accurate);
  • English → Whisper if its model is installed (word-level timestamps), otherwise SenseVoice locked to English.

One language per task — no mixed-language transcripts. The resolution is recorded in the task history (e.g. auto → sensevoice).

Manual pinning

Settings → Models → Transcription engine lets you pin SenseVoice or Whisper.

Also included

  • VAD chunking: Silero VAD pre-segments audio so only actual speech is processed — 30–50% faster on long videos;
  • Resume: interrupted tasks pick up where they stopped;
  • Re-transcribe: after switching engines or models, re-run any existing video with one click.

Released under the Apache-2.0 License