Subtitle Edit Whisper: Setup, Engines and Models Explained

Subtitle Edit Whisper support lives in one window: Video → Speech to text..., called Audio to text in older versions. Open a video, pick an engine and a model, choose the language and click Transcribe. Subtitle Edit downloads what it needs on first use. The real decision is which engine suits your hardware.

This post covers that choice for Windows with an NVIDIA card, Windows without one and Mac, the turbo model's one translation catch, batch mode, the quality report and what to check when nothing comes out. Facts about Subtitle Edit come from its GitHub repository and its official documentation, checked on 31 October 2026.

How do you use Whisper in Subtitle Edit?

Seven steps, taken from the speech to text documentation. Subtitle Edit is free and open source under the MIT licence; the current stable release is 5.3.0, from 14 October 2026.

1. Open a video file in Subtitle Edit.
2. Choose Video → Speech to text....
3. Select an Engine. Buttons beside it open the engine's website, download it and open its settings. Whisper CPP and Crisp ASR also show a Backend dropdown.
4. Select a Model. Larger usually means more accurate, slower and bigger on disk.
5. Select the Language of the audio, or auto-detect where the engine supports it.
6. Optionally tick Translate to English and Post-processing.
7. Click Transcribe.

Engines and models download automatically the first time you use them, so the first run includes a wait. The result is an ordinary subtitle in the main list, which you edit like any other file.

People still search for "subtitle edit audio to text". It is the same window under its old name, and a changelog line in the Subtitle Edit 5 series still uses it.

Which Whisper engine should you pick in Subtitle Edit?

It depends on your graphics card, because most engines in the list run the same Whisper models, packaged differently. The documentation's own advice: "For NVIDIA GPU users, use the Whisper CPP cuBLAS backend or Purfview Faster Whisper XXL for fastest transcription."

Purfview Faster Whisper XXL and Whisper Const-me are not available on macOS. Whisper OpenAI, the original Python implementation, is listed for every platform. The engine table in the docs is where each engine's platforms are kept current.

Other engines sit in the same list. Crisp ASR wraps several non-Whisper models, among them Parakeet, Canary and SenseVoice, and can detect speakers, marking lines with (speaker 0). Four online engines (OpenAI Compatible Server, OpenRouter, Alibaba Qwen3-ASR and Google Cloud Speech-to-Text) send your audio to a server, which matters if the footage is confidential. The rest run on your machine.

Subtitle Edit Whisper models: which one is right?

Pick the largest model your hardware finishes in a time you accept. The list runs from tiny, fastest and least accurate, through base, small and medium, to large, large-v2 and large-v3, the most accurate and the slowest.

A few specifics from the docs:

• .en models are English-only and better for English audio. Use them when every word is English.
• large-v3-turbo is listed as strong on Japanese. For most languages it is the sensible default when your hardware can run it.
• Language fine-tunes exist: .sv, .nb, and anime.ja, the anime-whisper model for anime and visual-novel Japanese.

Disk space and run time rise together. If you want numbers rather than rules of thumb, our Whisper model comparison lists the real download size of each model and what each one caught on a test clip. That is one clip, not a benchmark, but it shows what each step up costs.

Subtitle Edit Whisper GPU and CUDA: what actually helps

On Windows, a GPU means NVIDIA first. Three practical points:

1. Choose the backend, not just the engine. In Whisper CPP the Backend dropdown decides whether work runs on cuBLAS (NVIDIA CUDA), Vulkan or the CPU. On a machine with a good GPU, check it is not left on the CPU.
2. Match CUDA versions. Whisper CTranslate2 needs CUDA 12.x. If that engine gives you CUDA trouble, Whisper CPP cuBLAS and Purfview are the alternatives.
3. Out of memory? Go one model down. The documentation's answer to "CUDA out of memory" is a smaller model.

Two further settings sit under the engine's advanced options: per-engine command-line parameters, and VAD, which improves timing.

Translate to English returns the original language: the turbo catch

Translate to English converts non-English speech into English text in the same pass. Pick a turbo model, though, and the text can come back in the original language instead.

The reason is in OpenAI's Whisper README: the turbo model "is not trained for translation tasks" and returns the original language even when translation is requested. Its advice is to use medium or large. So if you tick the box with large-v3-turbo selected and see no English, change the model to medium or large and run it again.

That route only produces English. For other target languages, transcribe first, then use Translate → Auto-translate.... Its engines include Google Translate V1 (free), DeepL and ChatGPT (both with your own API key), LibreTranslate, MyMemory, and local models through Ollama, LM Studio or llama.cpp. Our guide to translating an SRT without breaking the timing covers what to check afterwards.

How do you transcribe many videos at once in Subtitle Edit?

Use Batch mode. Add your video files, choose the engine, model and language, and click Transcribe. Subtitle Edit writes an .srt file next to each video. Tick Add language code to file name if you want video.en.srt rather than video.srt, which matters as soon as several languages share one folder.

If your files vary a lot, such as a quiet interview next to a noisy street recording, run the hard ones as a separate batch with a larger model.

Post-processing and the transcription quality report

Whisper's output is accurate but untidy, and the Post-processing option tidies it as part of the transcription. The documentation lists what it can do:

• adjust timings, using waveform peaks
• fix short durations
• fix casing and add periods
• merge short lines and split long ones
• remove non-speech lines such as "[Music]"
• remove repeated lines

The transcription quality report is on by default, and it is the most useful part of the window for catching Whisper's habits. It flags five problems: too short, too long, overlap, non-speech and repeated. "Too long" includes what the docs call the typical hallucination, two or three words stretched over 10 to 30 seconds. "Repeated" means an engine stuck in a loop.

Treat every flagged cue as a place to listen, not just to delete. The patterns behind them, and the settings that prevent them, are in why Whisper repeats lines and invents text.

Subtitle Edit Whisper not working: no text, errors, CUDA

Start with the plain causes. Work down this list, which comes from the documentation's FAQ and tips:

1. Are the engine and the model downloaded? Subtitle Edit prompts for both. For an engine you have never used, let it download the engine and a model before you start.
2. Read the console log at the bottom of the Speech to text window. Right-click the window and choose View tools log file for the full record. Downloadable engines have a re-download item in the same menu.
3. Engine file gone missing? The usual cause is antivirus software quarantining the executable. Subtitle Edit checks when transcription starts, names the missing file and offers to download it again.
4. CUDA out of memory? Choose a smaller model.
5. Crisp ASR returns nothing with VAD on? Silero can reject a short clip as non-speech. Subtitle Edit retries once without VAD before it reports an empty result.

If the transcription finishes but the text is wrong rather than missing, check the language setting first, then try a larger model.

When does Sablate fit better than Subtitle Edit?

When the goal is captions ready to post rather than a subtitle file to refine, and you would rather not choose between engines and backends at all.

Sablate is a desktop app for Windows 10 or later and macOS 12 or later. It runs faster-whisper with four tiers: Fast (tiny), Balanced (small), Accurate (large-v3-turbo) and Best (large-v3). On Windows with an NVIDIA GPU, the CUDA runtime downloads on first GPU use; on a Mac it transcribes on the CPU. There is no engine or backend to choose.

What it adds is the second half of the job. One transcription feeds several translation layers, each with its own style and script font. Thirteen caption presets, including Hormozi and Karaoke, and a word-level karaoke sweep style the result, and burn-in handles 9:16, 1:1 or 16:9 by crop or blur pad. Word-by-word captions and multilingual subtitles show those workflows in detail.

The free plan needs no account and covers videos up to ten minutes, three jobs a month, the Fast and Balanced models and one offline translation layer; its burn-in carries a small corner mark. Pro costs $29 once or $5.99 a month, and the pricing page lists the rest. Other apps that work this way are compared in offline subtitle generators.

Subtitle Edit vs Sablate at a glance

Subtitle Edit wins on price, platforms, engine choice, speaker labels and format coverage. If you edit subtitles for a living or want full control of the pipeline, stay with it. Sablate is the shorter route to styled captions in several languages, for someone who would rather not read an engine table.

Which route should you take?

"I have an NVIDIA card and edit subtitles often." Stay in Subtitle Edit with Whisper CPP cuBLAS or Purfview Faster Whisper XXL, a large-v3-turbo model, and the quality report on.

"I have a laptop with no NVIDIA GPU." Whisper CPP with the Vulkan backend, a small or medium model, and patience on long files.

"I'm on a Mac." Whisper CPP first. Try WhisperX on Apple Silicon if you need word-level timing.

"The audio is not in English and I want English subtitles." Use medium or large with Translate to English, not turbo.

"Interviews with several speakers." Crisp ASR with speaker detection, which Sablate does not offer.

"I want vertical captions with word highlighting in three languages by tonight." Download Sablate for Windows or Mac.

Frequently asked questions

How do I use Whisper in Subtitle Edit?

Open a video, choose Video, then Speech to text, pick an engine and a model, set the spoken language and click Transcribe. Subtitle Edit downloads the engine and model the first time you use them. When it finishes, the lines appear in the main subtitle list, where you can edit them like any other subtitle.

Which Whisper engine is fastest in Subtitle Edit?

On an NVIDIA GPU, Whisper CPP with the cuBLAS backend or Purfview Faster Whisper XXL, which is what the official documentation recommends. Without NVIDIA, Whisper CPP with the Vulkan backend can use another GPU on Windows. On a Mac, Whisper CPP is the engine that runs everywhere, while WhisperX and CTranslate2 need Apple Silicon.

Which Whisper model should I choose in Subtitle Edit?

Choose large-v3-turbo if your computer can run it, small or medium if it cannot, and an .en model for English-only audio. Larger models are more accurate but slower and need more disk space. Tiny is fastest and least accurate, so keep it for quick drafts. Avoid turbo when you use Translate to English.

Why does Translate to English return the original language in Subtitle Edit?

Because turbo models are not trained for translation. OpenAI's Whisper documentation says a turbo model returns the original language even when translation is requested, and recommends medium or large instead. Switch the model in the Speech to text window to medium or large, run the transcription again, and the English translation should appear.

Why does Subtitle Edit Whisper give no text or an error?

Usually a missing engine or model, a file removed by antivirus, or too little GPU memory. Check that both are downloaded, read the console log at the bottom of the window, and choose a smaller model if CUDA runs out of memory. Subtitle Edit also names a missing engine file and offers to download it again.

Can Subtitle Edit transcribe many videos at once?

Yes. Batch mode lets you add several video files and click Transcribe, and it saves an SRT file next to each video. The option to add the language code to the file name gives names like video.en.srt. If one recording is much harder than the rest, run it as a separate batch with a larger model.