Subtitle Edit

the subtitle editor :)

Text to Speech

Generate speech audio from subtitle text using various TTS engines.

Text to Speech

How to Use

  1. Open Video → Text to speech…
  2. Select a TTS engine from the dropdown
  3. Select a language and voice
  4. Optionally enable Review audio segments to check each line before the final mix
  5. Optionally enable Add audio to video file to mux the result into a new video file (the settings button next to it opens the encoding settings); Advanced… below opens further TTS settings
  6. Click Generate speech from text to start
  7. Close the window with OK to apply the session’s subtitle changes (lines merged before generation, text edits made in the review window) to the subtitle in the main window — or Cancel to discard them

The bottom bar holds Set up cast (when speakers are present), Import…, OK, Cancel and Generate speech from text. OK is hidden while generating, and Cancel then stops the generation.

Remove the Original Speech From the Video

By default Add audio to video file replaces the video’s sound with the new speech, so music and sound effects are lost. Audio ducking (under Advanced…) keeps them by mixing the speech over the original track turned down - but the original voices are then still faintly audible under the new ones.

Remove original speech (slow) (also under Advanced…) gets rid of just the voices: it splits the video’s sound into speech and everything else, throws the speech away, and mixes the new speech over the music and sound effects. That is what a dub normally sounds like.

Set Up Cast: One Voice per Speaker

When the subtitle carries speaker names — the Actor field in ASSA/SSA, or <v Name> voices in WebVTT — a Set up cast button appears (with the speaker count). It opens a dialog where each actor is assigned an engine, voice, and optionally a model and voice instruction of their own. Lines without an actor use the globally selected voice. The cast is remembered between sessions, so the same actors open already assigned next time.

If the speaker names are written into the text instead (an SDH subtitle), see the speaker prompt below — and for a video where nobody labeled the speakers at all, Find Voices in Video and Clone works out the cast by listening.

Prompts Before Generation

Clicking Generate speech from text runs up to three quick checks on the subtitle before any audio is made. Each one only appears when it has something to show, each opens a review dialog where every proposed change is a checkbox, and each can be turned off in Options → Settings (search for “Text to speech: prompt”).

Speaker names in the text

An SDH subtitle writes its speakers into the text — MIKE: text, [NARRATOR] text, (Speaker 1) text, or a name alone on its line with the speech below. Sent to a TTS engine as-is, the names are read aloud. When no cast exists yet and at least two speakers are found, Subtitle Edit offers to move the names into the actor field: the tags leave the spoken text, the Set up cast dialog opens so each speaker gets a voice, and generation continues with the cast.

Merge continuation lines

Sentences split across several subtitles are offered for merging, so the engine speaks each thought as one breath group. These merges are applied to the subtitle in the main window when the window is closed with OK.

Skip sound and music lines

Lines that contain only sounds or music — ♪, [door slams], (sighs), or nothing once formatting tags are stripped — get read aloud or hallucinated into made-up words by TTS engines. Subtitle Edit offers to leave the checked lines silent: no audio is generated for them, and they are not counted as failures. A sound annotation followed by real speech ([gunshot] Get down!) is kept.

Speak From Current Line

Listen to the subtitle instead of generating an audio file: Speak from current line… reads the lines aloud one after another, starting at the selected line. It is meant for proofreading by ear - typos, awkward phrasing and bad line breaks are often easier to hear than to see.

  1. Select the line to start from
  2. Choose Speak from current line…
  3. Pick an engine, a voice and (for Supertonic) a language, then click OK - a missing engine or voice is downloaded at this point

Only the engines fast enough to keep up with speech are offered: Supertonic (the default), Kokoro TTS, Edge TTS and Piper (not on macOS). The choice is remembered for next time. The first line can take a few seconds while a local engine starts; after that the next line is generated while the current one plays, so the lines follow each other without gaps.

While reading, the selected line, the video position and the waveform follow the line being spoken. Time codes are ignored - each line is simply read after the previous one. The video is paused when reading starts.

Play With Speech

Watch a video with its subtitles spoken, like a live dub - handy for a course or a film in a language you do not speak, so you can watch the picture instead of reading. Play with speech… plays the video and speaks each line when the video reaches it. Nothing is saved; for an audio track to keep, use Text to speech above.

  1. Open the video and the (translated) subtitle, and move the video to where you want to start
  2. Choose Play with speech…
  3. Pick an engine, a voice and (for Supertonic) a language - the same fast engines as Speak from current line - and the options below, then click OK

The next few lines are generated ahead while the video plays. If the engine falls behind - for example while a local engine starts - the video waits for it. Pausing the video pauses the speech, seeking (clicking the waveform, a line or the position slider) continues from the new position, and the waveform toolbar’s Speed applies to both the video and the speech. The subtitle is read as it was when you started - edits made while playing are spoken after a restart.

Supported Engines

Local downloadable engines are installed into the Subtitle Edit data folder when you accept the download prompt.

CrispASR voice engines

Several of the local engines above are different models on the same CrispASR runtime, sharing one CrispASR/models folder with the speech-to-text backends. What separates them is voice output quality rather than features, so they are collected here.

Output rate is the engine’s native render rate - 48 kHz carries noticeably more high end than 24 kHz on headphones, though for speech mixed under a video the difference is small. Reference WAV is the format a cloning reference is converted to on import; Subtitle Edit resamples with ffmpeg for you, but a clean 3-10 second recording at or above that rate clones best.

Engine Output rate Languages Voice cloning Reference WAV Download
OmniVoice TTS (CrispASR) 24 kHz 646 Built-in voice + zero-shot 24 kHz mono ~1 - 1.6 GB
Qwen3 TTS (CrispASR) 24 kHz 10 VoiceDesign, CustomVoice or Voice clone 24 kHz mono (strictly enforced) ~2.3 GB
Chatterbox TTS (CrispASR) 24 kHz 23 on Base; Turbo is English-only Zero-shot 24 kHz mono ~700 MB - 1.8 GB Base, ~1 GB Turbo
IndexTTS (CrispASR) 24 kHz Follows the text Zero-shot 24 kHz mono ~600 MB - 2.4 GB
CosyVoice3 (CrispASR) 24 kHz 9, plus 18 Mandarin dialects as voices 8 baked-in presets + zero-shot 16 kHz mono + a transcript sidecar ~1.6 - 2.5 GB
MOSS-TTS (CrispASR) 24 kHz 20 Zero-shot 24 kHz mono ~10.5 - 20.5 GB incl. codec
Zonos TTS (CrispASR) 44.1 kHz 100+ via the language picker (trained on English, Japanese, Chinese, French and German; the rest rely on eSpeak pronunciation) None - one built-in voice - ~1.8 GB
VoxCPM2 (CrispASR) 48 kHz ~30 Zero-shot 24 kHz mono (upsampled internally) ~1.7 - 5 GB
dots.tts (CrispASR) 48 kHz Follows the text Zero-shot 24 kHz mono ~2.4 - 5 GB
VibeVoice (CrispASR) 24 kHz Follows the text Zero-shot 24 kHz mono ~1.6 - 5 GB
Confucius4-TTS (CrispASR) 22.05 kHz 14 Zero-shot (required - no default voice) 22.05 kHz mono ~1.9 - 2.6 GB
Pocket TTS (CrispASR) 24 kHz 6 (one model per language) Zero-shot, per request 24 kHz mono ~124 - 365 MB per language
Supertonic (CrispASR) 44.1 kHz 31 via the language picker None - 10 preset voices - ~200 MB

“Follows the text” means the engine has no language picker - it speaks whatever script it is given, taking its accent from the reference voice.

Notes on picking one:

Engine Settings

Some engines require additional configuration:

Voice Cloning

Several local engines can clone a voice from a reference recording. The first time you do this — when you import a reference recording, or generate speech with a cloned voice — Subtitle Edit shows a one-time dialog with the terms you are accepting. You have to tick the checkbox before you can continue.

The points that matter:

Declining just means “not now” — nothing is changed, the clone is refused, and you are asked again the next time. The answer is remembered per terms version, so you are asked again if the terms change.

Clean Voice-Clone References

Clean voice-clone references (Sidon) under Advanced… removes music, noise and reverb from the audio a voice is cloned from before the engine sees it: clips taken from the video (Clone voice to, Clone from video, Find voices in video and clone) and recordings imported in the voice manager. It uses the Sidon speech restoration model on audio.cpp (940 MB download, asked for on first use). The cloned voice no longer copies the background, but very quiet or overlapping speech can come out muffled. If the download is declined or cleaning fails, the reference is used as it is.

Renaming and Deleting Imported Voices

Right-click the voice combo box for Rename voice… and Delete voice…. Both work on voices you imported or cloned - the engines list those straight from their voices folder, one reference recording per voice - and are disabled for engine presets, built-in speakers and the Clone from video entry.

Supported for every file-backed cloning voice: the CrispASR engines (Chatterbox, Confucius4-TTS, CosyVoice3, dots.tts, IndexTTS, MOSS-TTS, OmniVoice, Pocket TTS, Qwen3 TTS, VibeVoice, VoxCPM2), the standalone OmniVoice TTS, and the audio.cpp engines (IndexTTS 2.5, Higgs Audio v3, Fish Audio S2 Pro, FireRedTTS3).

Voice Manager

Click the voice-manager button (the head-and-microphone icon next to the voice combo box) or pick Voice manager… from the combo box’s right-click menu to open one window for all voices of all engines.

Cloned voices are the only editable kind; an engine’s built-in presets are listed for orientation.

Cloning a Voice Heard in the Video

You do not have to prepare a reference recording by hand. In the main window, right-click a subtitle line in the waveform and choose Clone voice to → engine name. Subtitle Edit cuts the audio for that line out of the video’s current audio track, asks what to call the new voice (pre-filled with the line’s actor, or the video name and line number), and imports it as a cloned voice for the chosen engine. The line’s own text is used as the transcript the cloning engines want, so you are not asked to type it.

The menu item appears when exactly one subtitle line is selected and the right-click was on that line, and a video is open. Only engines that can clone are listed. The new voice is then in the voice list for that engine the next time you open Video → Text to speech….

Pick a line with clean speech: a couple of seconds or more, one speaker, and as little music and effects as possible. The clip is used exactly as it sounds in the video.

Find Voices in Video and Clone

Video → More → Find voices in video and clone… does the whole cast in one go: it works out who speaks in the video, clones each of them, and leaves the cast assigned so the dubbing is ready to generate.

What happens, in order:

  1. The video is transcribed with a speech-to-text engine that tells speakers apart — Crisp ASR MOSS Diarize is preselected (English and Chinese). You can change the engine in that window if you have another that labels speakers.
  2. The speaker labels the engine writes into the text ((Speaker 1) …) are moved into the subtitle’s Actor field, where they belong.
  3. A dialog lists the speakers with their line count, how much audio each has, and one of their lines so you can tell who is who. Give each speaker a real name — two speakers with the same name are merged into one voice, which is how you fix a person diarization split in two. Pick the cloning engine here as well.
  4. Each speaker’s voice is cloned from up to ~15 seconds of their own lines (their longest ones, joined), with those lines as the transcript.
  5. The subtitle switches to ASSA, which is the format that keeps actors, and the actor column is shown.
  6. The actor→voice cast is remembered, so Video → Text to speech… opens with every speaker already assigned to their cloned voice. Press Generate.

If a subtitle is already open, its lines and text are kept — the speakers are matched to your existing lines by time overlap, so a translation is not replaced by the transcription. With nothing open, the transcription becomes the subtitle.

Lines that overlap no detected speech (music, on-screen text) are left without an actor and fall back to the globally selected voice.

Clone From Video (Voice of Each Line)

For dubbing a video with several speakers there is a faster way than cloning each of them by hand. With a video open, pick Clone from video (voice of each line) at the top of the voice list. Every subtitle line is then spoken in the voice heard in the video at that line — whoever speaks line 12 in the original speaks line 12 in the dub. No reference recordings, no imports, no cast to assign.

Before generating, Subtitle Edit cuts one short reference clip per line out of the video’s audio. Lines shorter than about three seconds are grown into the silence around them, but never into the neighbouring line — a reference with two speakers in it would clone the wrong person.

Local SE5 Engines

Qwen3 TTS (CrispASR)

Qwen3 TTS runs through the CrispASR runtime and shares the CrispASR/models directory with the speech-to-text Crisp ASR engines.

Chatterbox TTS (CrispASR)

Chatterbox TTS runs through the CrispASR runtime (shared with the speech-to-text feature) and supports voice cloning.

Kokoro TTS

Kokoro TTS runs a local server with downloadable models.

OmniVoice TTS

OmniVoice TTS runs the omnivoice-tts CLI on CPU. It supports a large set of languages and voice cloning from reference WAV files (with an accompanying transcript). Because each line is a separate run that takes its own reference, it is also the engine behind Clone from video (voice of each line).

MistralSpeech

MistralSpeech is configured with an API key and model selection. The selected model is remembered in settings.

ElevenLabs Pauses and Pacing

You can control pausing and pacing in ElevenLabs output directly from the subtitle text.

Subtitle Edit sends your text to ElevenLabs with text normalization turned off, so these pause cues are preserved instead of being collapsed or rewritten.

The models offered are eleven_v4_turbo (the default), eleven_v4, eleven_turbo_v2_5, eleven_v3 and eleven_multilingual_v2. Behavior depends on the selected model:

Review Audio Clips

When Review audio segments is enabled, a dedicated review window opens after generation. This window lets you inspect, play, and regenerate audio for every subtitle line before the result is used. A 120px waveform of the original video audio is shown above the grid as a reference. The session (clips, per-line voice and engine, includes, text edits) can be written to SubtitleEditTts.json and imported again later — see Continuing a Session Later. Regeneration history is kept for the current session only and is not part of the export.

The Review Grid

Each subtitle line is shown as a row with the following columns:

Column Description
Enabled Checkbox to include or exclude the line from the final output
# Subtitle line number
Text The subtitle text (editable — double-click to modify before regenerating). Text edits are applied to the subtitle in the main window when the Text to speech window is closed with OK
Voice The voice used for that line
Speed The speed factor applied to fit the audio into the subtitle’s duration
CPS Characters per second for the subtitle line

Playback Controls

Right-click a line in the waveform for Play line, Regenerate audio, Show history and two timing fixes: Fit duration to generated audio ends the line where the generated speech ends, and Reset timing restores the times the line had when the window opened.

Regenerating a Clip

You can regenerate the audio for any individual line:

  1. Select the line in the grid
  2. Choose the desired engine, voice, language, model, or style from the dropdowns
  3. Click Regenerate or press R (or Ctrl+R)

The new clip is trimmed for silence and automatically speed-adjusted to fit the subtitle timing. After regeneration the new clip plays back immediately for review.

For ElevenLabs, extra fine-tuning parameters are available when that engine is selected: Stability, Similarity, Speaker Boost, Speed, and Style Exaggeration. A Reset button restores all ElevenLabs parameters to their defaults.

Regeneration History

Every time a clip is regenerated, the previous version is saved. To review the history for a line, click the History button on that row. The history dialog shows all generated versions with their voice name and speed, and lets you play each one. Selecting a version and clicking OK restores it as the active clip for that line.

Including / Excluding Lines

Uncheck the Enabled checkbox on any row to exclude that line’s audio from the final output. Excluded lines are skipped when the video file is assembled.

Exporting Clips

Click Export to save all audio clips and a SubtitleEditTts.json metadata file to a folder of your choice. The JSON file records the audio file names, subtitle timings, voice names, engine names, speed factors, and text for each line, making it easy to re-import or post-process the clips externally.

Continuing a Session Later

A long video does not have to be dubbed in one sitting. Export the session when you stop, and import it when you come back:

  1. In the review window, click Export… and pick a folder. Subtitle Edit writes SubtitleEditTts.json there, plus a wav subfolder with every generated clip.
  2. Next time, open Video → Text to speech… and click Import…, then pick that SubtitleEditTts.json.
  3. The review window opens again with all the lines and their audio, and you carry on where you left off.

What comes back with the session:

From there you can play, edit text, regenerate single lines, and finish the session normally. Keep the wav folder next to the JSON file: the clip paths are stored relative to it, so the whole folder can be moved or copied to another machine.

Keyboard Shortcuts

Key Action
R / Ctrl+R Regenerate selected line
Space Play / pause the selected line
Escape Close / Cancel
F1 Show help (default key, changeable)