@vexaai/transcript-rendering

Zero-dependency WebSocket transcript rendering pipeline: deduplication, speaker grouping, and timestamp parsing.

Used by Vexa and DNA.

Install

npm install @vexaai/transcript-rendering

Usage

import { deduplicateSegments, groupSegments, parseUTCTimestamp } from '@vexaai/transcript-rendering'

// Remove duplicate/overlapping segments
const clean = deduplicateSegments(segments)

// Group consecutive segments by speaker (default)
const groups = groupSegments(clean)

// Group by custom key
const groups = groupSegments(clean, {
  getGroupKey: (seg) => seg.language ?? 'unknown',
  maxCharsPerGroup: 1024,
})

API

deduplicateSegments<T>(segments: T[]): T[]

Removes duplicate and overlapping segments using multiple strategies:

• Adjacent duplicate detection
• Full containment — removes segments fully contained in another
• Expansion — merges partial segments into their completed versions
• Tail-repeat filtering — removes segments that are just repeated endings

Segments are scored by: known speaker (+10), completion status (+5), duration (0-3). Higher-scoring segments win conflicts.

groupSegments<T>(segments: T[], options?): SegmentGroup<T>[]

Groups consecutive segments by a configurable key (default: speaker). Options:

| Option | Default | Description | |--------|---------|-------------| | getGroupKey | seg => seg.speaker | Function returning the grouping key | | maxCharsPerGroup | 512 | Max characters before splitting a group |

parseUTCTimestamp(ts: string): number

Parses UTC timestamp strings (ISO 8601 or HH:MM:SS.mmm) into Unix epoch seconds.

Segment Interface

Your segments must have at minimum:

interface TranscriptSegment {
  text: string
  speaker?: string
  absolute_start_time: string
  absolute_end_time: string
  completed?: boolean
  start_time?: number
  end_time?: number
}

Extra fields pass through untouched via generics: deduplicateSegments<MySegment>(segments).

License

Apache-2.0