Making an AI podcast? The first step isn't comparing specs — it's clarifying your need: do you want a single voice reading a script, or a generated conversation between two hosts? These two categories run on completely different technology. Pick the wrong type and the effort is mostly wasted.
Solo narration is about text-to-speech: voice naturalness, rhythm, and stability over long passages. Multi-host dialogue is about conversation design: can listeners tell the speakers apart, are the transitions smooth, does the pacing feel like real people talking? If the tool doesn't match the job, no audio quality will save it.
First, figure out which of the four real needs is yours
- Solo narration: news, audio content, explainers — one voice from start to finish.
- AI-generated multi-host dialogue: turning documents into a two-host chat show.
- Recorded human dialogue: already recorded, needs transcription, editing, cleanup.
- Video podcast: publishing on video platforms too — needs visuals and captions.
Needs matrix
| What matters | Solo narration | Multi-host dialogue (AI-generated) |
|---|---|---|
| Core capability | TTS voice quality and long-form stability | Speaker separation and conversational rhythm |
| Chinese priority | High: rhythm, dialects, polyphonic characters | High: naturalness of Chinese dialogue |
| Biggest risk | Long passages turning robotic | Two voices sounding like one person acting |
| Post-production | Light — output is nearly finished | Medium — dialogue logic needs checking |
Tool comparison
| Tool | Best for | Chinese | Pricing | Positioning |
|---|---|---|---|---|
| NotebookLM | AI-generated multi-host dialogue | 50+ languages incl. Simplified Chinese | Free | Turn documents into a two-host chat show |
| ElevenLabs | Solo narration | 32 languages incl. Chinese | Free tier + pay per character | Natural voices with voice cloning |
| VibeVoice-1.5B | AI-generated multi-host dialogue | Test it yourself | Open source, self-hosted | Microsoft's open-source multi-speaker long-form voice |
| Descript | Post-production for recorded dialogue | English-first | Paid subscription | Text-driven editing — cut audio like a document |
| Chinese TTS services | Solo narration | Strong | Usage-based or subscription | Chinese rhythm and dialect voices are the strength |
Tool by tool
NotebookLM: the free two-host dialogue generator
From Google — upload documents and get an audio show with two AI hosts chatting, in 50+ languages including Simplified Chinese, for free. Great for turning reports and reading material into commute-friendly listening.
Not for: anyone who wants solo narration. It only does two-host dialogue — you can't pin down one fixed host voice. Also not for commercial projects that need a branded voice.
ElevenLabs: top-tier voices for solo narration
32 languages, a huge voice library, plus voice cloning — its emotional delivery and pauses in long-form narration are among the most natural in TTS today. The free tier (around 10,000 characters a month) is enough to test the waters.
Not for: anyone wanting one-click two-host dialogue. It's fundamentally a read-aloud tool; multi-role conversations must be stitched together manually. Long series burn through character credits fast — do the math on a tight budget.
VibeVoice-1.5B: a self-hosted multi-speaker option
Microsoft's open-source podcast-grade speech synthesis generates up to 90 minutes of four-speaker conversation in one go — a fit for technical teams building their own pipeline. Microsoft's open-source VibeVoice-1.5B multi-speaker podcast voice (/article/132)
Not for: ordinary users who don't want to deal with deployment. Compute, tuning, and Chinese quality checks are all on you — the time cost is real.
Descript: the editing workhorse for recorded dialogue
For human-recorded multi-host conversations, the pain is post-production, not generation: transcription, filler-word removal, multi-track alignment. Descript turns audio into editable text — deleting a sentence deletes that audio segment. Descript's text-driven podcast editing (/article/663)
Not for: anyone trying to generate a podcast from nothing. It doesn't create from scratch — useless without recordings. Chinese transcription lags behind English.
Chinese solo narration: prioritize Chinese rhythm
For pure Chinese solo shows, domestic TTS services are worth trying first: polyphonic characters, number reading, and dialect voices usually come out smoother. When evaluating, listen to pauses and stress in long paragraphs — not just the first three sentences of the demo.
Not for: anyone needing multilingual versions. Chinese TTS engines generally pronounce English less naturally than native multilingual models.
Three selection rules
- Lock the show format first, then pick the tool. Solo narration means TTS, multi-host dialogue means conversation design, recorded shows mean editing tools. Wrong format, everything after it is wrong.
- Test Chinese shows with Chinese scripts. The gap in Chinese dialogue naturalness between tools is huge — generate three minutes from your own script in each and compare; don't trust official demos.
- Start free, pay once the workflow works. Free tiers are enough to validate a workflow — subscribe only after the publishing cadence is sustainable.
Alternative: if nothing fully satisfies you, a hybrid stack is often more reliable — Chinese TTS for solo parts, transcribed and polished recordings for dialogue. This guide covers transcription and caption tool selection (/article/1862). To also publish on video platforms, just add a video-editing step.
FAQ
Q: With zero budget, can I still make a decent AI podcast?
A: Yes. NotebookLM generates two-host dialogue for free, and ElevenLabs' free tier covers a few short episodes. Validate your direction first, get the rhythm going, then think about paying.
Q: Is AI-generated Chinese multi-host dialogue actually usable?
A: A notch below English. Speaker separation and conversational rhythm give themselves away more easily in Chinese — test with a three-minute sample before committing to a series.
Q: Can I mix tools for solo and dialogue segments?
A: Yes, but don't expect one tool to do everything. Common pattern: TTS for solo parts, separately generated or human-recorded dialogue, then assemble everything in editing software.
