fix: stop feeding running ASR output back as Whisper initial prompt (repetition loops)

[NaturalPsi]

Problem

Whisper can fall into a repetition loop: once it emits a phrase, that phrase reappears every window until the audio ends. In our use this wiped out large stretches of real speech — a 56‑minute Japanese recording had minutes 20–47 replaced entirely by はいはいはい…‑style loops, even though the audio was clear and well above the noise floor (the loop region measured −25 dB RMS vs −24 dB in a clean region, so it was not a volume problem).

The trigger is self‑conditioning. generateInitialPrompt() (local Whisper) and the Amical Cloud request both feed the running ASR output (previousTranscription / currentAggregatedTranscription) back in as the next window's prompt. When a hallucinated phrase slips in, re‑injecting it as context makes Whisper far more likely to emit it again, which re‑injects it again — a self‑reinforcing loop. whisper.cpp's own no_context default is true precisely to avoid this; the app re‑introduces the carry‑over at the application layer.

Change

Stop feeding the running transcription back as the initial prompt, in both providers:

• whisper-provider.ts – generateInitialPrompt() no longer passes previousTranscription.
• amical-cloud-provider.ts – the request sends previousTranscription: undefined.

Vocabulary and cursor context (beforeText) are preserved — those are static and don't loop, so spelling/term hints still work.

Verification

Built from this branch (local Whisper, large‑v3, Metal) and re‑transcribed the same recording: the previously‑lost minutes came back as clean Japanese with no repetition loops (~5.8× more real text recovered). Equivalent to running whisper.cpp with no_context=true, which is its default.

Tradeoff / open question

This is the minimal fix and disables cross‑window context entirely. There's a small coherence cost (the model no longer "remembers" the last few words across a long dictation). If you'd prefer to keep some context, happy to rework this as:

• a setting (conditionOnPreviousText, default off), or
• a guarded injection — only carry context when the previous window wasn't flagged as low‑confidence / a detected n‑gram repeat.

Let me know which direction you'd prefer and I'll adjust.

Summary by CodeRabbit

Bug Fixes

• Updated transcription processing to remove context carry-over between transcription calls, eliminating self-reinforcement behavior in audio transcription.

[coderabbitai]

📝 Walkthrough

Walkthrough

Both the Amical Cloud and Whisper transcription providers remove prior aggregated transcription context from their request and prompt building logic. The previousTranscription field in cloud requests and prompt generation is now explicitly undefined instead of passing the running transcription output, disabling the self-reinforcement loop.

Changes

Remove prior transcription context from providers

Layer / File(s)	Summary
Disable prior transcription context in cloud and Whisper providers apps/desktop/src/pipeline/providers/transcription/amical-cloud-provider.ts, apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts	AmicalCloudProvider's HTTP /transcribe request and WhisperProvider's initial prompt generation both stop passing prior aggregated transcription as context. The previousTranscription field is now explicitly undefined with added inline comments, and unused aggregatedTranscription parameters are marked as such.

🎯 2 (Simple) | ⏱️ ~8 minutes

A fluffy fix to silence the echo, no more loops of old transcription—
Two providers now speak fresh, unhindered by what came before! 🐰✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: stopping the re-injection of running ASR output as initial prompt to fix repetition loops in Whisper, which directly addresses the core problem and changes in both provider files.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

🧹 Nitpick comments (1)

apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts (1)

284-297: ⚡ Quick win

Remove the now-unused aggregatedTranscription parameter from the prompt helper signature.

The behavior is correct, but keeping an intentionally-unused argument here leaves stale API intent and adds noise for future edits.

Suggested cleanup diff

-  private generateInitialPrompt(
-    vocabulary?: readonly string[],
-    aggregatedTranscription?: string,
-    accessibilityContext?: TranscribeContext["accessibilityContext"],
-  ): string {
+  private generateInitialPrompt(
+    vocabulary?: readonly string[],
+    accessibilityContext?: TranscribeContext["accessibilityContext"],
+  ): string {
@@
-    void aggregatedTranscription;
     const prompt = buildWhisperPrompt({
       vocabulary,
       previousTranscription: undefined,
       beforeText:
         accessibilityContext?.context?.textSelection?.preSelectionText,
     });

       const initialPrompt = this.generateInitialPrompt(
         context.vocabulary,
-        aggregatedTranscription,
         context.accessibilityContext,
       );

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts` around
lines 284 - 297, The generateInitialPrompt function currently declares an unused
aggregatedTranscription parameter and explicitly voids it; remove
aggregatedTranscription from the parameter list and delete the `void
aggregatedTranscription;` statement inside generateInitialPrompt (function name:
generateInitialPrompt) and update any call sites to stop passing that argument
so the signature and usage remain consistent with buildWhisperPrompt's expected
inputs; ensure TypeScript types compile after removing the parameter.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts`:
- Around line 284-297: The generateInitialPrompt function currently declares an
unused aggregatedTranscription parameter and explicitly voids it; remove
aggregatedTranscription from the parameter list and delete the `void
aggregatedTranscription;` statement inside generateInitialPrompt (function name:
generateInitialPrompt) and update any call sites to stop passing that argument
so the signature and usage remain consistent with buildWhisperPrompt's expected
inputs; ensure TypeScript types compile after removing the parameter.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 7fc7c6c0-855c-4059-a726-162525c2c4ce

📥 Commits

Reviewing files that changed from the base of the PR and between 9219e87 and 5b17639.

📒 Files selected for processing (2)

• apps/desktop/src/pipeline/providers/transcription/amical-cloud-provider.ts
• apps/desktop/src/pipeline/providers/transcription/whisper-provider.ts

[haritabh-z01]

Hey @NaturalPsi thanks for flagging this.
Would you be able to run a few checks with only last x tokens being fed (besides the vocabulary)
I will try running longer tests / multiple tests over the coming weekend.