MMS Audio-Text Aligner (SRT)

The fastest way to turn "I have the transcript" into "I have the timestamps"

Most subtitle workflows in ComfyUI go one direction: Whisper hears audio, guesses the text, and timestamps its own guess. This node goes the other way. You already have the transcript - a script, translated lyrics, captions that drifted, a dubbed line - and you want to know exactly when each word is spoken. That's forced alignment, and this is the rare node that does it in the graph and hands you a ready-to-use .srt.

It's the difference between someone summarizing your audio and someone placing a ruler on it. If your text and audio say the same thing, you get word-level start/end times, and from those you build subtitles. Perfect for karaoke-style captions, translated subs, or re-timing a Whisper transcript that's close but off by half a second.

How it works

Under the hood it's Meta's facebook/mms-1b-fl102, a 1B-parameter speech model that covers 102 languages (trained on the cleaner FLEURS dataset). Each language gets its own adapter weights and native tokenizer - no romanization needed for most scripts.

The pipeline: your audio is resampled to 16 kHz mono in memory, the transcript's language is detected with lingua (when language=auto), the model runs one forward pass to get CTC emissions, and torchaudio.functional.forced_align does the actual alignment math. Token spans get merged into word boundaries, then grouped into SRT cues. All in RAM - no temp WAV files dumped on your disk.

The big catch is upfront, not in the pipeline: the ~3.2 GB model downloads to ComfyUI/models/mms/facebook__mms-1b-fl102/ on your first run, not at install. Plan for a slow first execution; every run after uses the cache.

The inputs that matter

The node has a lot of knobs, but a beginner really sets these:

• audio - from any node that outputs ComfyUI's AUDIO type, normally LoadAudio.
• text - the transcript, multiline. Punctuation and newlines here are usable as segmentation cues.
• language - auto detects from the text. If detection fails, pick your language from the dropdown or enter an ISO 639-3 code via custom.
• segmentation_mode - single_word gives one word per cue (karaoke), max_chars+punctuation+newlines is the sensible default for reading subs.
• gap_ms - bridges sub-300 ms silences between cues so single-word subtitles don't flicker. Set 0 to disable.
• output_name - base filename, no extension; .srt is appended automatically.

Everything else is optional tuning: max_chars_per_line, remove_punctuation (keeps apostrophes and hyphens, so "don't" survives), split_count to split into name_1.srt, name_2.srt…, romanize as a fallback, and precision (bf16 default - fp8 is experimental and wants PyTorch 2.1+).

Two outputs: srt (the full subtitle text, always populated, wire it to a Preview Text node to check without writing a file) and output_path (newline-separated list of files written to ComfyUI's output folder).

Installing it

ComfyUI Manager: search "MMS-Aligner". Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/hnvcam/ComfyUI-MMS-Aligner
cd ComfyUI-MMS-Aligner
pip install -r requirements.txt

Then restart and the node appears under audio/alignment as "MMS Audio-Text Aligner (SRT)". torch and torchaudio come from ComfyUI; transformers, huggingface_hub, uroman, and lingua-language-detector are pure Python - no compiler, no MSVC, which is exactly what the author promises. One genuine gotcha: the README's own install snippet still says idealweek/ComfyUI-MMS-Aligner (an earlier repo name - the live remote is hnvcam). If you copy-paste the stale URL, you'll be chasing a repo that moved. Clone the hnvcam one.

Also note faster-whisper in requirements.txt is commented as optional - you only need it for chunk_with_whisper=True, which splits long audio (roughly 15–20+ minutes on consumer GPUs) after MMS OOMs on it.

Troubleshooting

• "Could not auto-detect language" - your text is too short or mixed. Just set language explicitly.
• "Transcript produced no alignable tokens" - the text has characters outside the language's tokenizer vocab. Try romanize=True, or double-check your ISO 639-3 code.
• Slow first run - it's the 3.2 GB download. Not a bug; it caches.
• Your language isn't in FL102's 102 languages - the README is blunt that this node won't work for you. Check before you sink a download into it.
• Out of memory on long clips - drop precision to fp16, or turn on chunk_with_whisper so a rough Whisper pass finds safe split points and each chunk aligns independently.

If you've got a clean transcript and you're tired of Whisper's timestamps drifting off the audio, this is the node you're actually looking for.