📺 TTS SRT

This is the node you reach for when the audio has to land on the subtitles, not just near them. Feed it an SRT file and a voice, and it generates speech that fits each cue's start and end time - dubbing a video, narrating to an existing timeline, making a voiceover that hits its marks. If you just want to read a block of text aloud with no timing constraints, its sibling 🎤 TTS Text is the simpler pick. TTS SRT is the one that cares about the clock.

The important thing to understand up front: this node doesn't contain a voice model. It's a consumer. You wire an engine node (VibeVoice, Higgs Audio, ChatterBox, F5-TTS, Qwen3-TTS, whatever) into its TTS_engine input, and TTS SRT drives that engine cue by cue. Same node, any engine - that's the whole "unified" idea of the suite.

How it works

You give it srt_content - literally the text of a .srt file, timestamps and all. The node parses each subtitle block, generates that line through your chosen engine, then reconciles the generated length against the cue's allotted window. That reconciliation is the timing_mode, and it's the setting that actually changes your results:

• smart_natural (the default) shifts timing intelligently to avoid overlaps while keeping speech sounding natural. Start here.
• stretch_to_fit time-stretches audio to exactly match each cue. Precise, but push it too far and voices get chipmunky or draggy.
• pad_with_silence generates naturally and pads gaps with silence.
• concatenate ignores the timestamps and just plays the lines back to back.

Character and language switching work inside the SRT text too - drop [Alice] or [de:Bob] into a cue and it switches voice or language for that line, same syntax as the rest of the suite.

The inputs and outputs that matter

• TTS_engine - required. This is the engine node's output. Nothing happens without it.
• srt_content - your subtitles as text. The default shows the expected format, including [Alice]/[Bob] character tags.
• narrator_voice - the default voice, picked from the bundled voices_examples/ list (Attenborough, Eastwood, male/female refs, and more). Or leave it none and drive the voice from an engine that clones.
• timing_mode and seed round out the required set.

For fine control there's max_stretch_ratio / min_stretch_ratio (caps on how far stretch_to_fit will warp), timing_tolerance, enable_audio_cache (leave it on - it only regenerates cues you changed), and use_native_duration_targeting for engines like OmniVoice that can aim for a duration at generation time.

Four outputs. audio is the finished track - wire it to a Preview/Save Audio node. Adjusted_SRT is the gold one people miss: an SRT with the actual generated timings, so your subtitles and audio stay in sync downstream. generation_info and timing_report are text diagnostics - read timing_report when a cue sounds rushed and it'll tell you which lines it had to stretch.

Installing it

It ships with the whole pack. Via ComfyUI Manager, search "TTS Audio Suite", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

Run install.py in your ComfyUI's own Python env, then restart. The engine models download themselves on first use.

Common issues & troubleshooting

"Nothing generates." You almost certainly forgot the TTS_engine connection, or it's connected but that engine's model hasn't finished its first-run download. Check the console.

Every line sounds sped-up or slowed-down. That's stretch_to_fit fighting cues that are too short for the words. Widen min_stretch_ratio/max_stretch_ratio to give it room, or switch to smart_natural and let it shift timing instead of warping pitch. Genuinely tight cues just need shorter text.

Voices drift or cut off on long scripts. Timing quality is inherited from the engine you plugged in - a long-form engine like VibeVoice handles extended SRTs better than a short-context one. Check timing_report to see where it struggled.