fix: retry empty prompted transcriptions

[jatinkrmalik]

Summary

• retry once without custom vocabulary when a prompted transcription returns no usable text
• preserve vocabulary prompting for successful decodes and explicit language prefill during fallback
• cover the retry condition with a focused unit test

Root cause

WhisperKit can exit during promptTokens prefill before decoding audio on affected models, returning an empty transcript. Retrying without custom prompt tokens recovers the dictation while keeping vocabulary bias on the normal path.

Upstream context: argmaxinc/argmax-oss-swift#372 and argmaxinc/argmax-oss-swift#497.

Validation

• swift test (225 passed, 1 environment-dependent test skipped, 0 failures)

Closes #181

[jatinkrmalik]

/build

[github-actions]

⏳ PR Build started for fa1bed6

Build signed & notarized DMG... this usually takes 10–20 minutes.

Watch the build →

[github-actions]

✅ PR Build ready!

DMG	VocaMac-0.7.0-pr.182+fa1bed6-arm64.dmg
Size	5.1M
Branch	jmalik/fix-whisper-prompt-fallback
Commit	fa1bed6
Signed	✅ Developer ID
Notarized	✅ Apple

📥 Install

1. Click the DMG link above to download
2. Open the DMG and drag VocaMac to Applications (replace existing)
3. Open VocaMac — No Gatekeeper warnings, no permission resets

SHA-256 checksum

21a63fdca79b299e2bf37c658f5ec12ba56e59e17a4d07931e572a7f867ba10d  VocaMac-0.7.0-pr.182+fa1bed6-arm64.dmg

💡 Comment /build or /build-quick to rebuild.

[jatinkrmalik]

@flesler, the signed and notarized test build is ready: download the DMG. Could you try it with Large v3 Turbo (Compact) and custom vocabulary enabled, using the same audio that reproduced the empty transcript? Let us know whether it returns text.

[flesler]

Tested 0.7.0-pr.182+fa1bed6 on M5 Max with Large v3 Turbo (Compact) and custom vocabulary restored (GraphQL, HTML, Git, AWS, MongoDB, Zod, Jira, UTF-8, JSON).

Results

#	duration	prompted	after retry	notes
1	6.2s	empty → retry	“Okay, this is a text about MongoDB running on AWS and using GraphQL.”	✅
2	6.6s	empty → retry	“This is another one… MongoDB… JSON and HTML.”	✅
3	4.6s	empty → retry	“A normal sentence without jargon…”	✅
4	3.0s	empty → retry	“one more with git, html and sod”	✅ (Zod→sod, expected without vocab bias)
5	0.7s	empty → retry → still empty	blank	❌ silent PCM (#176), not vocab

Every real dictation logged:

Prompted transcription was empty for openai_whisper-large-v3-v20240930_turbo_632MB; retrying without custom vocabulary

then returned usable text. Vocab mitigation works.

Take 5 is the separate silent-capture bug from #176/#178: buffer was near-silent, so retrying without prompt tokens correctly still yields nothing. #182 isn’t meant to fix that.

Note

On the retry path vocabulary bias is gone — “Zod” → “sod”. Fine for unblocking; custom vocab still isn’t actually helping on this turbo model until upstream prompt prefill is fixed.

I’d call #181 verified for this build on the vocab path.