Speech to text for Record mode. While you record, Subtitld streams your microphone to it: the words show as you speak, and each finished sentence fills its subtitle. It runs faster-whisper models on your own computer, downloaded once from Hugging Face.

What it does

Words as you speak

In Record mode the text appears while you talk and is refined as you go. Subtitld shows it on its own until the sentence is finished.

Sentences in place

Each finished sentence carries the time it was said, so it fills the subtitle cut at that moment.

Keeps your last words

When you pause, it waits for the sentence you were still saying, up to a minute, and adds it before it finishes.

Warm between takes

From one take to the next in the same language the model stays loaded, so later takes start without loading it again.

To use it, pick it as the Transcription engine in the Audio tab, under Recording.

In the transcription panel, use Range or Selection. Subtitld cuts those to WAV, the only audio this add-on reads. With All it gets the video's own audio and stops with bad_audio. Each run gives one subtitle for the whole stretch, so it suits filling in a selected subtitle best. A run can't be cancelled: it always finishes.

Models

Pick the model in Model: tiny, base, small (the default), medium or large-v3, and the English-only .en versions of all but large-v3. Bigger models are more accurate, slower and heavier.

The model downloads from Hugging Face the first time it is used, and is kept for next time. Loading it takes from seconds to a minute for the large ones. In Record mode it is loaded twice, once for the live text and once for the finished sentences, so if your computer runs short of memory, pick a smaller one.

Options

In Settings › Add-ons › RealtimeSTT › Configure.

Model

Default: small

The Whisper model; see Models above. Bigger is more accurate, slower and heavier. Names ending in .en understand only English.

Device

Default: auto

Where the model runs. In the builds Subtitld installs, auto always picks the CPU: it asks PyTorch for a CUDA GPU, and those builds carry PyTorch without CUDA. Choose cuda to ask for an NVIDIA GPU directly; CI doesn't test that.

Compute type

Default: default

The precision of the numbers the model works with. default lets the engine choose; int8 is fastest on a CPU; float16 suits GPUs.

Stream interim results

Default: on

In the transcription panel, show text while a phrase is still being decoded; the final text replaces it. Record mode always shows live text, whatever this says.

The add-on reads these when it starts, so a change applies the next time it starts, for example after you restart Subtitld.

Languages

Regional tags go to Whisper as their main language: pt-br is sent as pt.

en
pt
pt-br
es
fr
de
it
nl
ru
pl
uk
tr
ar
zh
ja
ko
hi
id
sv
fi
da
no
cs
el
he
ro
hu
ca
th
vi

Platforms

System	Download	Tested
Linux x86-64	482.7 MB	In CI, without a model
macOS (Apple silicon)	377.0 MB	In CI, without a model
Windows x86-64	280.6 MB	In CI, without a model

Before each release, CI runs the add-on's tests on every system and starts the built add-on to check that its speech detection loads. No transcription runs in CI, since that needs a model.

Troubleshooting

bad_audioThe audio isn't WAV. In the transcription panel, use Range or Selection instead of All.
internalTranscription failed; the message says why. On first use the model may not have downloaded: check your connection and try again. If you set Device to cuda, go back to auto.
not_installedThe speech engine inside the add-on didn't load. Reinstall the add-on.

License

Add-on

MIT

Engine

MIT

RealtimeSTT, faster-whisper

Models

MIT

OpenAI Whisper weights

Fine for commercial work: the add-on, RealtimeSTT, faster-whisper and the Whisper weights are all MIT.

RealtimeSTT

What it does

Models

Options

Languages

Platforms

Troubleshooting

License

Versions

Other Speech to text add-ons