Skip to content

PRD-008: Automatic Speaker Name Detection

  • Status: ✅ Implemented (v2.1.0)
  • Related RFCs: RFC-010

Summary

Automatically detect and label speaker names (hosts and guests) in Whisper transcriptions using Named Entity Recognition (NER), eliminating the need for manual --speaker-names configuration.

Background & Context

When Whisper generates transcriptions, speakers are labeled generically as "Host" and "Guest" unless the user manually provides names via --speaker-names. This manual step:

  • Requires feed-specific knowledge that users may not have
  • Breaks when episodes have unique guests
  • Adds friction when running at scale or embedding the pipeline
  • Reduces transcript quality and readability

Users want transcripts that automatically use real speaker names without requiring manual configuration for each podcast or episode.

Goals

  • Automatically extract host names from RSS feed metadata (author tags, titles, descriptions)
  • Automatically detect guest names from episode metadata (titles, descriptions)
  • Provide speaker names to Whisper screenplay formatting for higher-quality transcripts
  • Maintain backward compatibility with manual --speaker-names as fallback
  • Support multiple languages through configurable NER models

Non-Goals

  • Voice fingerprinting or audio-based speaker identification
  • Multi-speaker diarization beyond host/guest distinction
  • Real-time speaker detection during transcription

User Stories

  1. As a podcast archivist, I want transcripts to automatically show "Lenny Rachitsky" and "guest name" instead of "Host" and "Guest" so my archive is more searchable and readable.

  2. As a researcher, I want to process hundreds of episodes without manually looking up speaker names for each one.

  3. As a developer integrating the pipeline, I want automatic speaker detection so I don't need to maintain a database of podcast hosts.

Functional Requirements

Host Detection (Feed-Level)

  1. Extract host names from RSS author tags (highest priority):
  2. RSS 2.0 <author> tag
  3. iTunes <itunes:author> tag

  4. iTunes <itunes:owner><itunes:name> tag

  5. Fall back to NER extraction from feed title/description if no author tags

  6. Validate detected hosts by checking if they appear in episode content

Guest Detection (Episode-Level)

  1. Extract guest names from episode titles (e.g., "Interview with John Smith")
  2. Extract from episode descriptions using NER
  3. Filter out detected hosts to avoid duplication

Configuration

Option Default Description
auto_speakers true Enable automatic speaker detection
language "en" Language for NER model selection
ner_model derived from language Override NER model (e.g., en_core_web_sm)
speaker_names [] Manual fallback names if detection fails

CLI Flags

```bash

Enable/disable auto detection

--auto-speakers / --no-auto-speakers

Set language (affects both NER and Whisper)

--language en

Override NER model

--ner-model en_core_web_sm

Manual fallback

--speaker-names "Lenny" "Guest" ```python

Success Criteria

  • ✅ Host names correctly detected from RSS author metadata
  • ✅ Guest names correctly extracted from episode titles/descriptions
  • ✅ Detected names appear in Whisper screenplay output
  • ✅ Manual --speaker-names works as fallback when detection fails
  • ✅ No regression in transcript quality when names aren't detected
  • ✅ Performance impact < 100ms per episode for NER processing

Dependencies

  • spaCy: NER library for named entity extraction
  • spaCy language models: en_core_web_sm (English), others for additional languages