Forced-alignment in Subtitle Edit via CrispASR

[subof]

Hi,
I decided to put this here, as I believe that a joint discussion will help to find a solution as soon as possible. Forced alignment is a fundamental task, since most ready-made transcripts are very inaccurate with respect to timestamps. The Montreal Forced Aligner (MFA) console utility is a standard tool in such cases, but it's not a very convenient way for most users. In addition, CrispASR opens up new LLM-based features that are more visual and modern.
In Subtitle Edit (SE), this function is implemented via the File/Import/Plain text menu. However, all attempts to carry out this process have failed. In the current version, I enter the command:
--max-len 50 --vad --vad-model "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\ggml-silero-v6.2.0.bin" --align-only --aligner-model canary-ctc-aligner-q8_0.gguf
I added --align-only as a flag indicating that this is not a new transcription process.
log:

Date: 07/28/2026 12:34:47
SE: v5.1.0-rc18 - Microsoft Windows NT 10.0.19045.0 - 64-bit
Calling speech-to-text (Crisp ASR Canary) with : C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\crispasr.exe --backend canary -l en -m "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\models\canary-1b-v2-q8_0.gguf" -am "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\models\canary-ctc-aligner-q8_0.gguf" -f "C:\Users\USer\AppData\Local\Temp\se-stt-76f1b533-970b-43e0-92b4-8dac93fa3b1e\a3a1e8e3-bdce-49d6-93c4-9d8581c9b9bd.wav" --output-srt --print-progress --max-len 50 --vad --vad-model "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\ggml-silero-v6.2.0.bin"  --align-only --aligner-model canary-ctc-aligner-q8_0.gguf
crispasr 0.8.23 (git 7d22deec, Release) [backends: cpu,cuda]
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 24563 MiB):
Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6, VMM: yes, VRAM: 24563 MiB
crispasr[align-only]: requires --ref-text or --text-file.
Speech to text (Crisp ASR Canary) done in 00:00:00.3137007
Loading result from STDOUT

In other cases, I did not add the --align-only flag, but the process still could not complete successfully and the subtitles turned out to be much less accurate, actually spoiling. There is no ALIGN button inside the settings, and command selection turns into a quest. I think it should be intuitive.

[subof]

I downloaded the VAD model manually, because it didn't work through the program. But it didn't change anything.
ggml-silero-v6.2.0.bin

Calling speech-to-text (Crisp ASR Canary) with : C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\crispasr.exe --backend canary -l en -m "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\models\canary-1b-v2-q8_0.gguf" -am "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\models\wav2vec2-xlsr-en-q4_k.gguf" -f "C:\Users\User\AppData\Local\Temp\se-stt-8fc6df10-2d4d-4287-9d72-6176834474b6\5da05266-b019-4ac1-bd41-0072f8e4573c.wav" --output-srt --print-progress --max-len 50 --split-on-punct --vad --vad-model "C:\Users\User\AppData\Roaming\Subtitle Edit\CrispASR\ggml-silero-v6.2.0.bin"  --align-only --aligner-model canary-ctc-aligner-q8_0.gguf
crispasr 0.8.23 (git 7d22deec, Release) [backends: cpu,cuda]
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 24563 MiB):
Device 0: NVIDIA GeForce RTX 3090 Ti, compute capability 8.6, VMM: yes, VRAM: 24563 MiB
crispasr[align-only]: requires --ref-text or --text-file.
Speech to text (Crisp ASR Canary) done in 00:00:09.0118891

[AppleSheeple]

Yeah. I couldn't make the --align-only CLI interface workable either.

With --align-granularity word, you get single word transcripts.

And I couldn't find a way to actually save results.

The input srt support doesn't do what you may think it does. It's just parsed as text from what I saw.

I actually ended up using the rust crate to get what I want, which was turning JSON output without words to JSON output with words. And that was easy by calling align_words() on each transcript separately with the right pcm samples and t_offset.

Ideally for the CLI, one can do each step separately with JSON -> JSON. We have that with VAD now, but not with align. And I'm guessing not with diarize either, although I didn't try that yet.

And the last step would be going from JSON to one of the subtitle formats, which I believe the CLI currently also lacks.

But maybe the programmable interface is a better fit for my ideals in any case.

[CrispStrobe]

Forced alignment itself works — the gap seems to be that Subtitle Edit never passes the transcript, so there is nothing for --align-only to align. crispasr[align-only]: requires --ref-text or --text-file is correct behaviour (exit 10), not a failure of the aligner.

Reproduced and verified end-to-end on the v0.8.24 release binary with canary-ctc-aligner:

crispasr -f audio.wav -am canary-ctc-aligner-q8_0.gguf \
         --align-only --text-file transcript.txt \
         --align-granularity segment --output-srt

On samples/jfk.wav with a 3-line transcript that returns exactly what you want — one cue per input line, on stdout, which is where SE already reads from:

1
00:00:00,400 --> 00:00:02,160
And so my fellow Americans

2
00:00:03,520 --> 00:00:07,280
ask not what your country can do for you

Re-timing an existing .srt works too, and is probably the more useful path for you: pass the .srt as --text-file. It is detected by extension, cue text is kept, timestamps/indices are discarded, and one cue in gives one re-timed cue out — no --align-granularity needed. I fed it a deliberately wrong file (all cues at 0–1s, 1–2s, 2–3s) and got the correct times back.

Two problems in the command line from the log, worth fixing regardless:

• --aligner-model canary-ctc-aligner-q8_0.gguf in the custom-arguments box is an alias for -am, so it overrode the full path SE had already passed with a bare relative filename. Drop it — SE's -am is enough.
• --backend canary -m canary-1b-v2-q8_0.gguf loads the 1B ASR model, which --align-only never uses. Alignment needs only -am. Dropping it saves the load entirely.

--max-len and --vad likewise belong to the transcription path, not to alignment.

So nothing here needs a CrispASR change to work today — but I agree it should not be a quest. The clean fix is SE-side and small: when Import Plain text runs against a CrispASR engine, write the imported text to a temp .txt (exactly as SE already writes the temp .wav) and pass --text-file <that> plus --align-only. That is the one piece of information only SE has.

Happy to add whatever makes that easier on our side — e.g. accepting the transcript on stdin so no temp file is needed, or a clearer error that names --text-file and points at the aligner docs. @niksedk Say which would suit SE best and I will ship it.

[niksedk]

We have it in a PR here: SubtitleEdit/subtitleedit#12833

Will ship with v5.2.0 .
v5.1.0 is language frozen (and mostly feature frozen).