BATS - Speech2Subtitle

Subtitles from spoken speech – Whisper with word-exact timing, in three hardware variants

Speech2Subtitle turns an audio file (or the audio of a clip) into an SRT file with exact word timing – using Whisper, the proven AI transcription model. The “Add function...” assistant offers three variants, one per hardware: Intel-optimized (OpenVINO: NPU/GPU/CPU), Nvidia-optimized (Whisper-Faster) or iNPU-optimized (Whisper-NPU for Intel NPUs). All three transcribe in ~90 languages, can optionally translate the result to English, and file the finished SRT directly in the BIN.

Adding the menu entry

In the assistant (right-click in the BIN → Add function...) pick the fitting variant – Speech2Subtitle - Intel optimized, … Nvidia optimized or … iNPU optimized. The shared questions:

1. Language: “Auto-Detect The Language” or one of ~90 languages (Afrikaans … Yoruba).
2. Translate? “Do you want to translate the result to English?” – Yes makes Whisper transcribe directly into English text (e.g. for foreign-language interviews).
3. Scope: the entire video or the clip’s IN/OUT part.
4. Placement (BIN or timeline) and name of the menu entry – the assistants pre-fill it sensibly (e.g. “German, standard, Model Medium, NPU”).

Variant-specific questions on top:

• Intel (OpenVINO): device (Intel NPU / Intel GPU / CPU – AMD included), subtitle format (standard: max. 42 characters per line, max. two lines · standard asia: max. 16 characters per line · one sentence per line – great for TTS · one word per line) and model size (Tiny, Base, Small, Medium – recommended, Large V2, Large V3; bigger = more precise, but slower).
• Nvidia (Whisper-Faster): the same subtitle format, plus high precision (float16) – mandatory from the RTX 5000 series up, simply “No” without an Nvidia GPU – and model size (Tiny … Large V3, Turbo; Medium recommended).
• iNPU (Whisper-NPU): choose the hardware (NPU / GPU / CPU – must be Intel hardware) and the subtitle format. The model is fixed at medium here, and a VAD phase detection (silence detection) is permanently active. Note from the assistant: keep the Intel NPU driver current (version 32.0.100.3053 or newer).

How to use it

1. Right-click an audio clip in the BIN → Generate Subtitle submenu → your entry (e.g. “German, standard, Model Medium, NPU”).
2. BAT Server first extracts the audio (16 kHz mono, optionally the IN/OUT part) and lets Whisper transcribe it – progress in the job display.
3. The finished SRT is registered in the BIN under __BAT Server\Subtitle automatically (file stamp _stt, with a format suffix for the Intel variant). From there it can be taken into an MKV via SubtitleMerge, or used as a speech source for Text2Speech/Convert2Speech.

Good to know

• Variant = hardware: Nvidia GPU → Nvidia variant (fastest on RTX); Intel system with NPU → iNPU variant (and a current NPU driver); everything else (Intel/AMD without NPU) → Intel OpenVINO variant (GPU or CPU).
• Weigh the model size: Medium is the sensible default; Large/Turbo are more precise (esp. filler words, technical terms) but take considerably longer – with the Nvidia variant, float16 is mandatory from RTX 5000 up.
• Line formats: “standard”/“standard asia” are meant for display as subtitles; “one sentence per line” optimizes the SRT for downstream Text2Speech presets, “one word per line” for maximum granularity (e.g. marker/tag workflows).
• New file, original kept: only the SRT file is created; the audio original remains untouched.

Managing the preset

Entries created via the assistant are user presets: the .bat0 file lives under C:\ProgramData\EdiusPowerTools\BAT-server\bat_files\BIN\ or …\bat_files\timeline\ depending on your choice, and can be freely edited, renamed or deleted (deleting removes the context-menu entry). Language, translation, format, model and device are baked into the .bat0 – changes take effect from the next run.