perf: switch podcast TTS to multi-speaker single request

[logosc]

Problem

The podcast agent currently generates audio by making one TTS API call per script segment sequentially. A typical podcast has ~47 segments, each taking ~8s → ~6 minutes total for audio generation.

Benchmark: 47 segments, 9.27MB MP3, 5m59s generation time.

Solution

Gemini TTS supports multiSpeakerVoiceConfig which allows two speakers in a single request. Since our podcast uses exactly two hosts (A and B), the entire script can be sent as one conversation.

Tested with a 6-segment sample:

• 1 API call instead of 47
• 10.6s response time → 21s MP3 with two distinct voices (Kore + Charon)

Request format

{
  "generationConfig": {
    "responseModalities": ["AUDIO"],
    "speechConfig": {
      "multiSpeakerVoiceConfig": {
        "speakerVoiceConfigs": [
          {"speaker": "A", "voiceConfig": {"prebuiltVoiceConfig": {"voiceName": "Kore"}}},
          {"speaker": "B", "voiceConfig": {"prebuiltVoiceConfig": {"voiceName": "Charon"}}}
        ]
      }
    }
  }
}

Implementation plan

1. Add generateMultiSpeakerTTS() — formats segments as A: text\nB: text\n..., uses multiSpeakerVoiceConfig
2. Chunking fallback — if script exceeds ~8000 chars, split into segment groups and run up to 4 parallel workers
3. Replace the sequential for-loop in GenerateAudioTool.Execute
4. Remove old single-voice generateTTS() function

Expected result

Audio generation drops from ~6min to <30s for a typical podcast.