Sound-Effects Models Ranked (2026): Text-to-SFX and Video-to-SFX Together

Do you need a sound from a sentence, or a sound that follows a picture? Those are different tasks. This page puts our twelve text-to-SFX routes and four video-to-SFX routes in one score-ordered view, with all 128 rated clips playable. It keeps every original score and cost basis, and labels the two cross-mode family averages as summaries rather than new measurements.

By Syver-August, Developer, North Noir · October 27, 2026 · 8 min read

One view, two tasks, no invented common test

Our text-only benchmark contains 96 blind ratings from twelve delivered routes, recorded on 25 October 2026. Our video-to-SFX benchmark adds 32 ratings from four routes; its completed rating records were supplied on 27 October. Syver-August rated both sets. The eight cue names match, but the models were not given the same kind of input.

Text and video scores are not interchangeable: different input, the same eight cue names and the same rater. Text routes received text prompts with a four-second target (native duration exceptions retained); video routes received ten-second silent videos. These were separate blind packs, not a head-to-head preference test. Rank uses each route-task row’s own unrounded score; it does not measure whether adding video improves a model.

The 16 rows are route-task rows, not sixteen distinct model families. A route tested on text has a text row; a route tested on video has a video row. There is no score for an unrun task and no zero-filled cell to make the table look complete. Tango 2 delivered no text clips inside that run’s queue deadline, so it has no listening score and is excluded from this ranked set.

Every measured row still uses mean rating × 10 and equal cue weights. ±SEM is sample SD across eight cue ratings divided by √8, then multiplied by 10. It is not a listener-consensus measure or a confidence interval. Cost and timing can change bar width, never bar height or rank. The text dataset does not publish comparable end-to-end timings, so those widths say N/A rather than implying zero.

What the ordering says, and what it cannot say

Stable Audio Open’s 75.00% text score and Mirelo SFX 2.0’s 73.75% text score sit above Sonilo v1.1’s 72.50% video score. That is the ordering of the recorded means, not evidence that text input is better than video input. A text prompt asks a model to imagine a cue. A video asks it to follow a particular visible event, with its own timing and ambiguity. The ten-second video also gives more room for an unwanted voice or music event than a four-second text target.

The gallery and notes make this distinction useful. In the video pack Sonilo’s gravel scored 8/10 for gravel steps only; Mirelo’s 6/10 note said it sounded like the man rather than the gravel. In the text pack Mirelo’s gravel scored 8/10, “Matches the requested sound very well.” Sonilo’s scored 7/10, “Sounds correct.” These are cue-specific observations from different outputs and routes, not a controlled demonstration that video made one model better or worse.

Two family rollups, with the arithmetic left visible

Cross-mode averages are summaries, not new measurements or extra ranked rows. Each is (text Quality % + video Quality %) / 2, only for a family tested in both modes. Sonilo uses v1.1 in both; Mirelo compares text SFX 2.0 with video SFX v1.5, not the same version. Per-mode scores stay visible. No pooled SEM, significance test or score for an unrun mode is inferred.

Sonilo: (70.00% text + 72.50% video) / 2 = 71.25%. Mirelo: (73.75% text + 61.25% video) / 2 = 67.50%. Sonilo is v1.1 in both modes. Mirelo’s text route is SFX 2.0; its video route is SFX v1.5. A family label does not erase that version difference. The rollups remain separate from the sixteen measured rows and do not acquire an invented pooled SEM.

One ranked list, with every clip in place

The ranked list above follows the chart’s order across all sixteen route-task rows, without separate text and video galleries. Every row carries its eight rated clips: the original native audio files for text, or the exact ten-second video muxes with generated sound for video. The four-second text target and native duration exceptions remain visible. No files, scores or evidence URLs have changed. Notes were lightly edited for spelling and grammar; meaning and scores are unchanged. Start at low volume; levels were not matched.

Compare request costs without pretending the outputs match

The table imports every cost from its original data module. Text targets four seconds, with native duration exceptions; video inputs are ten seconds. Mirelo video requires two samples and its cost includes both, even though sample 0 alone was rated. Compute-priced routes use returned meters rather than output length. These are tariff-derived figures, not verified invoices, and no source-video creation cost is hidden inside the audio request cost.

That makes the cost-width switch a browsing aid, not a fair equal-duration price race. A low bill does not compensate for an unwanted event in the quality score, and a higher bill is not evidence of a better cue. The original articles preserve each request’s batch, duration and tariff basis so you can compare the task you actually need.

Verdict: choose the task, then audition the evidence

For text-only effects, Stable Audio Open led this listener’s test; for sound from a picture, Sonilo v1.1 led the video pack. Sonilo also has the higher of the two cross-mode family summaries. None of those statements says a single family wins every task, that a small gap is statistically established, or that another listener would agree. Use the table to shortlist, then play the cue in the mode you intend to use.

The full companion articles are Text-to-SFX Models Ranked and Video-to-SFX Models Ranked. They remain separate evidence sets with their original slugs and scope. For an AI-drawn film, the useful result is a sound you want to keep against the picture, not an average you have to defend.

Key takeaways

• 16 measured route-task rows: twelve text and four video, not sixteen model families. Each row retains the original score, SEM and cost basis.
• All 96 text clips and all 32 video-with-audio clips are playable. Same cue names and rater, different input and duration: scores are not interchangeable.
• Sonilo’s cross-mode family average is 71.25%; Mirelo’s is 67.50%. These are arithmetic summaries, not new measurements or extra ranked rows.
• Mirelo’s text and video routes use different versions. No unrun mode, missing timing or delivery failure is assigned a fake zero.

Frequently asked questions

Can text-to-SFX and video-to-SFX scores be compared directly?

Not as interchangeable measurements. The same listener rated the same eight cue names, but text models received text prompts with a four-second target and video models received ten-second silent pictures. The combined table orders their own recorded means, not a shared-input preference test.

Why are there sixteen rows but only two family averages?

The table has twelve measured text routes and four measured video routes. Only Sonilo and Mirelo families were tested in both modes, so only they receive cross-mode summaries. No missing mode is invented or scored zero.

How are the cross-mode averages calculated?

Sonilo: (70.00 + 72.50) / 2 = 71.25%. Mirelo: (73.75 + 61.25) / 2 = 67.50%. Both per-mode numbers remain visible. These are summaries, not new measurements; Mirelo compares text SFX 2.0 with video SFX v1.5.

Can I hear the evidence behind both sets of scores?

Yes. All 96 native text outputs and all 32 exact rated video muxes are playable on this page, with scores and listening notes. Video files contain the picture plus generated sound, not audio-only substitutes.

Explore AI-drawn video

Keep reading

• Text-to-SFX Models Ranked (2026): Blind Listening and Prices
• Video-to-SFX Models Ranked (2026): Four Models on Eight Real Clips
• Background Music Models Ranked (2026): Vocals, Instrumental and Prices