Podcast Generator

Step 2: Format Selection

Choose the podcast format:

Step 3: Script Writing — NotebookLM Pattern

The two-host format that works (reverse-engineered from Google's Audio Overviews):

• Host A = "The Explainer" — knows the material, breaks down concepts
• Host B = "The Questioner" — audience surrogate, asks the "wait, why?" questions
• Dialog rhythm: alternate short punchy lines with longer explanations. Sprinkle affirmations: "Right.", "Exactly.", "Okay so—"
• Arc: open with common misconception → introduce source that challenges it → unpack implications → "so what does this mean for you"
• Transitions: "And on that note..." / "Which brings us to..." / "Here's where it gets weird—"

Structure (target ~150 words per minute of audio):

1. Cold open (15–30s) — the single most surprising finding, stated as a question or contradiction
2. Setup (30–60s) — what we're covering, why it matters now
3. Segments (3–5 × 2–4 min) — one idea each; end each with a mini-hook into the next
4. Takeaways (1–2 min) — 3 things to remember
5. Outro (15s) — sign-off

Write for ears, not eyes: contractions always, no semicolons, no parentheticals. If you wouldn't say it out loud, rewrite it.

Script format — one line per utterance, speaker tag in brackets, blank line between speakers. This is the unit you'll chunk for TTS:

[ALEX]: So today we're diving into something that honestly broke my brain a little.

[SAM]: Oh no. What now.

[ALEX]: Okay — you know how everyone says [common belief]? There's this paper from [source] that basically says... the opposite.

[SAM]: Wait. The *opposite* opposite?

Step 4: Audio Generation — ElevenLabs

Install: pip install elevenlabs pydub

Model choice: eleven_multilingual_v2 for quality (10K char limit per call); eleven_turbo_v2_5 for speed/cost (40K char limit, ~300ms latency, ~3x faster).

Voice IDs that work for duo podcasts (from the default library — verify with client.voices.search()):

• JBFqnCBsd6RMkjVDRZzb (George — warm, mid-range male)
• 21m00Tcm4TlvDq8ikWAM (Rachel — clear, measured female)
• pNInz6obpgDQGcFmaJgB (Adam — energetic narrator)
• EXAVITQu4vr4xnSDxMaL (Bella — conversational female)

Settings for conversational podcast delivery:

• stability: 0.45 — lower = more expressive; below 0.3 gets inconsistent
• similarity_boost: 0.8 — keeps voice consistent across chunks
• style: 0.3 — mild exaggeration for energy (0 = flat)
• use_speaker_boost: True

import os
from elevenlabs.client import ElevenLabs
from pydub import AudioSegment
import io

client = ElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])
VOICES = {"ALEX": "JBFqnCBsd6RMkjVDRZzb", "SAM": "21m00Tcm4TlvDq8ikWAM"}

def render_line(speaker: str, text: str) -> AudioSegment:
    audio = client.text_to_speech.convert(
        voice_id=VOICES[speaker],
        text=text,
        model_id="eleven_multilingual_v2",
        output_format="mp3_44100_128",
        voice_settings={"stability": 0.45, "similarity_boost": 0.8,
                        "style": 0.3, "use_speaker_boost": True},
    )
    return AudioSegment.from_mp3(io.BytesIO(b"".join(audio)))

# parse script → list of (speaker, text) tuples, render each, concat
gap = AudioSegment.silent(duration=350)  # 350ms between speakers
episode = sum((render_line(s, t) + gap for s, t in lines), AudioSegment.empty())
episode.export("episode_raw.mp3", format="mp3", bitrate="128k")

Chunking long utterances: split at sentence boundaries (., ?, !), keep under ~800 chars per call. Pass previous_text/next_text params to preserve prosody across chunk boundaries.

Step 5: Loudness Normalization

Podcast standard is -16 LUFS (stereo) per Apple/Spotify specs. pydub's normalize() is peak-only — not LUFS. Use ffmpeg's two-pass loudnorm via the ffmpeg-normalize wrapper:

pip install ffmpeg-normalize
ffmpeg-normalize episode_raw.mp3 -o episode.mp3 -c:a libmp3lame -b:a 128k \
    -t -16 -tp -1.5 -lra 11 --normalization-type ebu

-t -16 = target LUFS, -tp -1.5 = true-peak ceiling (prevents clipping), -lra 11 = loudness range. This runs two passes automatically (analyze, then correct).

Episode Length Guidelines

Rule of thumb: ~150 words per minute of audio.

Best Practices

1. Hook early — if the first 30 seconds aren't interesting, listeners skip
2. One idea per segment — don't cram too much; let ideas breathe
3. Use stories and examples — abstract concepts need concrete illustrations
4. Vary pacing — alternate between fast energy and slow, thoughtful moments
5. End with value — give listeners a clear takeaway or action item

Limitations

• Requires ELEVENLABS_API_KEY env var
• Voices mispronounce technical terms/acronyms — spell phonetically in the script ("Kubernetes" → "koo-ber-NET-eez") or use ElevenLabs pronunciation dictionaries
• eleven_multilingual_v2 has known issues with very long single calls (voice drift, occasional stutter) — chunk at sentence boundaries, don't send 5K-char blobs
• Cost: ~$0.18–0.30 per 1000 characters depending on plan; a 20-min episode (~3000 words) ≈ $3–5