Podcast Generation

1. Install components:
   • pip install kokoro-onnx (TTS — fast, CPU-friendly)
   • ollama pull llama3.1:8b (script generation)
2. Pipeline: LLM generates a dialogue script → each speaker's lines synthesized with different Kokoro voices → audio segments concatenated with slight gaps.

import ollama, soundfile as sf
from kokoro_onnx import Kokoro

kokoro = Kokoro("kokoro-v0_19.onnx", "voices.json")

# Step 1: Generate dialogue script
resp = ollama.chat(model="llama3.1:8b", messages=[{
    "role": "user",
    "content": "Write a 2-person podcast script about local AI: host 'Sarah' and guest 'Mike'. Topic: running AI models on your own computer. 3-minute script. Format: SARAH: <text>, MIKE: <text>."
}])
script = resp["message"]["content"]

# Step 2: Parse and synthesize each line
audio_segments = []
for line in script.split("\n"):
    if line.startswith("SARAH:"):
        samples, sr = kokoro.create(line[7:], voice="af_sarah")
    elif line.startswith("MIKE:"):
        samples, sr = kokoro.create(line[7:], voice="am_michael")
    audio_segments.append(samples)

# Step 3: Concatenate with 0.5s silence gaps
import numpy as np
silence = np.zeros(int(0.5 * sr))
final = np.concatenate([np.concatenate([seg, silence]) for seg in audio_segments])
sf.write("podcast.wav", final, sr)

1. A 3-minute podcast generates in 1-3 minutes on CPU.

Podcast generation is CPU-friendly. Kokoro TTS + Llama 3.2 3B (script writing) runs entirely on CPU. A $300 laptop generates a 10-minute podcast in 5-10 minutes. The LLM script generation is the bottleneck — 3B models write slower but for structured scripts (host+guest dialogue), quality is adequate. For higher-quality scripts: add a used GTX 1060 6 GB ($60) for the LLM at 40-60 tok/s. Total: ~$360. Podcast generation at $300-400 produces credible AI podcasts with distinct voices, natural pacing, and coherent dialogue. The output is good enough for internal communications, educational content, and hobby projects.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs Llama 3.1 8B for high-quality script writing at 50-80 tok/s + Kokoro TTS for voice synthesis. A 30-minute podcast generates in 5-10 minutes. For production podcast pipelines (daily AI-generated news summaries, internal comms): batch generate overnight — 8 hours = ~50 episodes. Total build: ~$700-900. For voice variety: Kokoro has ~20 preset voices. For cloned voices (CEO/host voice cloning), F5-TTS adds VRAM overhead (+4-6 GB) but yields personalized podcasts.

The mistake: Generating a 20-minute AI podcast script, running it through TTS with default settings, and publishing without listening to the whole thing. Why it fails: AI-generated scripts hallucinate facts and statistics. TTS mispronounces technical terms and names (Kokoro reads "Qwen" as "kwen," "Ollama" as "oh-la-ma"). A 20-minute podcast with 5 factual errors and 10 mispronunciations destroys credibility instantly. The fix: Always do a human listen-through before publishing. Fact-check every claim the LLM makes. Add pronunciation guides to the script: spell "Qwen" as "Queen" and "Ollama" as "Oh-la-ma." Edit the script for natural speech — written text ≠ spoken text. AI generates the first draft; human polish makes it publishable. An AI-generated podcast with obvious AI errors gets one thing: ignored.