feat: fill timing gaps step

CLAUDE.md

@@ -70,6 +70,7 @@ podwhisperer/
 │       ├── index.ts             # Lambda handler (durable workflow)
 │       ├── types.ts             # Shared types
 │       ├── steps/               # Pipeline step implementations
+│       │   ├── fill-timing-gaps.ts
 │       │   ├── llm-refinement.ts
 │       │   ├── replacement.ts
 │       │   └── segments-normalization.ts
@@ -100,7 +101,7 @@ podwhisperer/
 2. **Pipeline Lambda** (durable execution) orchestrates the workflow:
    - Sends job to SQS queue
    - Waits for callback from GPU worker
-   - Applies post-processing steps (replacement, LLM, normalization)
+   - Applies post-processing steps (fill timing gaps, replacement, LLM, normalization)
    - Generates captions (VTT, SRT, JSON)
    - Sends EventBridge notification
 3. **GPU Worker** (ECS container on Managed Instances):
@@ -124,7 +125,7 @@ podwhisperer/
 
 **Configuration** (`packages/config/index.ts`):
 - Zod schemas for all pipeline configuration
-- Exported types: `PipelineConfig`, `TranscriptionConfig`, etc.
+- Exported types: `PipelineConfig`, `TranscriptionConfig`, `FillTimingGapsConfig`, etc.
 - Used by both CDK (at synth time) and Lambda (at runtime)
 
 ## Important Implementation Notes

README.md

@@ -39,6 +39,7 @@ flowchart TD
         CheckExists{Output Exists?}
         SQS[Send to SQS]
         WaitCallback[Wait for Callback]
+        FillGaps[Fill Timing Gaps]
         Replacement[Replacement Rules]
         LLM[LLM Refinement]
         Normalize[Segments Normalization]
@@ -67,10 +68,10 @@ flowchart TD
 
     S3Upload --> EB --> CheckExists
     CheckExists -->|No| SQS --> Poll
-    CheckExists -->|Yes| Replacement
+    CheckExists -->|Yes| FillGaps
     Poll --> Download --> Transcribe --> Align --> Diarize --> Upload --> Callback
     Upload --> Raw
-    Callback --> WaitCallback --> Replacement --> LLM --> Normalize --> Finalize --> Captions --> Notify
+    Callback --> WaitCallback --> FillGaps --> Replacement --> LLM --> Normalize --> Finalize --> Captions --> Notify
     Finalize --> Refined
     Captions --> VTT & SRT & JSON
 ```
@@ -88,7 +89,15 @@ The WhisperX container runs on GPU-enabled ECS Managed Instances:
 5. **Diarization** - Identifies and labels different speakers
 6. **Upload** - Saves raw transcript to S3 `output/` prefix
 
-### Step 1: Replacement Rules (Optional)
+### Step 1: Fill Timing Gaps
+
+WhisperX sometimes produces words with missing `start`/`end` timestamps, particularly for words containing numbers or currency symbols (e.g., `"15"`, `"$8"`, `"$0.25"`). This step fills those gaps using character-proportional interpolation between neighboring anchor words, giving all downstream steps clean timing data.
+
+- Gaps are distributed proportionally based on word character length
+- Padding is applied dynamically based on the segment's speech rate
+- Filled words are marked with `score: 0` to distinguish them from original timestamps
+
+### Step 2: Replacement Rules (Optional)
 
 Apply regex or literal string replacements to fix common transcription errors:
 
@@ -101,23 +110,23 @@ Apply regex or literal string replacements to fix common transcription errors:
 }
 ```
 
-### Step 2: LLM Refinement (Optional)
+### Step 3: LLM Refinement (Optional)
 
 Uses Amazon Bedrock to improve transcript quality:
 
 - Fixes spelling and grammar errors while preserving meaning
 - Identifies speakers by name when possible (e.g., "SPEAKER_00" becomes "Luciano")
 - Validates suggestions to prevent aggressive rewrites
 
-### Step 3: Segments Normalization
+### Step 4: Segments Normalization
 
 Splits long transcript segments into caption-friendly chunks:
 
 - Respects character and word limits per segment
 - Splits at natural punctuation boundaries
 - Handles speaker changes within segments
 
-### Step 4: Caption Generation
+### Step 5: Caption Generation
 
 Generates multiple caption formats from the refined transcript:
 
@@ -129,7 +138,7 @@ Optional features:
 - Word-by-word highlighting (underline, bold, or italic)
 - Speaker name prefixes (always, when speaker changes, or never)
 
-### Step 5: Notification
+### Step 6: Notification
 
 Sends an EventBridge event when the pipeline completes, including:
 
@@ -270,6 +279,12 @@ The pipeline will automatically trigger and generate outputs in `s3://$BUCKET/ou
 | `skipIfOutputExists` | boolean | `false` | Skip transcription if output already exists |
 | `hfTokenSsmPath` | string | `"/podwhisperer/hf_token"` | SSM parameter path for HuggingFace token |
 
+### Fill Timing Gaps
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `enabled` | boolean | `true` | Enable fill-timing-gaps step |
+
 ### Replacement Rules
 
 Array of replacement rules, each with:

lambdas/pipeline/index.ts

@@ -20,6 +20,7 @@ import {
 import { SendMessageCommand, SQSClient } from '@aws-sdk/client-sqs'
 import { PipelineConfigSchema } from '@podwhisperer/config'
 import type { EventBridgeEvent } from 'aws-lambda'
+import { fillTimingGaps } from './steps/fill-timing-gaps'
 import { llmRefinement } from './steps/llm-refinement'
 import { applyReplacements } from './steps/replacement'
 import { normalizeSegments } from './steps/segments-normalization'
@@ -167,6 +168,42 @@ const handler = async (event: S3EventBridgeEvent, context: DurableContext) => {
 
   // refining steps
 
+  // Fill timing gaps — fills missing word timestamps before any other processing
+  await context.step('fill-timing-gaps', async () => {
+    if (!pipelineConfig.fillTimingGaps.enabled) {
+      context.logger.info('Fill timing gaps disabled, skipping step')
+      return
+    }
+
+    // Load the transcript from S3
+    const response = await s3.send(
+      new GetObjectCommand({
+        Bucket: BUCKET_NAME,
+        Key: transcriptProcessingKey,
+      }),
+    )
+    const bodyStr = await response.Body?.transformToString()
+    if (!bodyStr) {
+      throw new Error('Empty response body from S3')
+    }
+    const transcript = JSON.parse(bodyStr) as WhisperxResult
+
+    // Fill timing gaps
+    const stats = fillTimingGaps(transcript, pipelineConfig.fillTimingGaps)
+
+    // Save the updated transcript back to S3
+    await s3.send(
+      new PutObjectCommand({
+        Bucket: BUCKET_NAME,
+        Key: transcriptProcessingKey,
+        Body: JSON.stringify(transcript),
+        ContentType: 'application/json',
+      }),
+    )
+
+    context.logger.info('Filled word timing gaps', { stats })
+  })
+
   // Replacement rules
   await context.step('replacement-rules', async () => {
     if (typeof pipelineConfig.replacementRules === 'undefined') {