Blog

Last month I decided to launch a weekly podcast for my side project. Problem: I hate the sound of my own recorded voice. I sound nasally, I trip over words, and editing out “ums” takes me longer than writing the script.

So I went looking for an AI voice solution — something that could turn my written scripts into a clean, natural-sounding podcast without sounding like a robot from 2015.

I spent two weeks testing. Burned through 7 tools. Got frustrated. Almost gave up.

Here’s what I learned, what I built, and the stack I’m actually using today.

The Problem with Most “AI Voice” Articles

Most guides online tell you to use ElevenLabs or Murf — both great products, but at $22-99/month they’re overkill for a hobby project. The free tiers give you 10-30 minutes of audio per month. That’s one episode.

So I set a constraint: zero recurring cost. One-time purchases or free only. If I needed to buy a tool outright, fine — but no subscriptions.

What I Wanted

• A podcast-quality narrator voice (not “Siri reading your script”)
• Support for at least English, ideally multiple languages
• Runs on my Windows laptop (no cloud uploads)
• Under $50 total setup cost
• Easy enough that I don’t need to re-learn it every week

Tool #1: Piper TTS — Where I Started

Piper is a local neural TTS engine. It’s lightweight (~50MB per voice model), fast (<1s per sentence on CPU), and supports 140+ languages. You can install it with pip install piper-tts and be running in 5 minutes.

First impression: It works. The voices are clean and crisp. But they sound… read. Like someone doing a professional audiobook narration, not someone having a conversation.

The open-source community has trained dozens of voice models, but most are optimized for clarity, not natural conversation.

What I learned: Piper is great for announcements, tutorials, and anything where a “neutral narrator” voice works. It’s not great for conversational podcast segments.

Verdict: Keep it in the toolbox for intros/outros, but not my main voice.

Image credit: Unsplash

Tool #2: Edge-TTS (Microsoft) — The Free King

Edge-TTS is a Python library that taps into Microsoft’s neural TTS engine — the same one powering Edge’s “Read Aloud” feature. Install with pip install edge-tts.

First impression: Holy crap, this sounds real.

The default “en-US-JennyNeural” voice has natural intonation, proper pauses, and even breath sounds. It’s not perfect — sometimes the emphasis lands on the wrong word — but it’s shockingly good for something that costs $0.

The catch: It requires internet. And there’s a rate limit — if you try to generate 20 minutes of audio at once, it’ll throttle you.

My workaround: I generate audio in 3-minute chunks and concatenate them with FFmpeg. Takes a few extra seconds but avoids throttling entirely.

import edge_tts
import asyncio

async def generate_chunk(text, filename, voice="en-US-JennyNeural"):
    communicate = edge_tts.Communicate(text, voice)
    await communicate.save(filename)

# Split your script into 3-min chunks
chunks = [script[i:i+2000] for i in range(0, len(script), 2000)]
for i, chunk in enumerate(chunks):
    asyncio.run(generate_chunk(chunk, f"chunk_{i}.mp3"))

Then concatenate:

ffmpeg -f concat -safe 0 -i file_list.txt -c copy final_episode.mp3

Verdict: My primary podcast voice. Free, high quality, and the chunking workaround is trivial.

Tool #3: AudioCraft (Meta) — Music & SFX Generation

Once I had the voice sorted, I needed intro music and background tracks. Meta’s AudioCraft can generate music from text prompts, entirely locally.

Installation is a bit involved (requires Conda + PyTorch), but once it’s running, you can generate 30-second music clips from prompts like “upbeat electronic intro, 120 BPM, synthwave style”.

Reality check: The output is impressive for an open-source model, but it’s not going to replace a professional composer. The music has that “AI-generated” feel — slightly washed out, sometimes incoherent.

Best use case: Background ambient tracks where the music isn’t the focus. I use it for my podcast’s intro sting and transition sounds.

The Final Stack

Here’s what I’m running now, total cost:

Total: $0. Everything is free. The only investment was my time learning the tools.

What a Real Episode Sounds Like

I published Episode 1 of my test podcast last week. Here’s the workflow:

1. Write script in plain text (~1500 words per 10-min episode)
2. Generate voice with Edge-TTS in 3-min chunks (~2 minutes total generation)
3. Generate intro with AudioCraft (~30 seconds, 2-3 retries)
4. Stitch in Audacity — voice track + intro + fade outs
5. Export MP3 and upload to Anchor

Total time from script to published episode: about 45 minutes. Google’s NotebookLM podcast feature takes 5 minutes but gives you zero control. This takes longer but I get to edit, re-record specific segments, and keep full ownership.

Lessons Learned

What Works

• Edge-TTS for voice: Hands down the best free option for natural-sounding narration
• Chunking scripts: Avoids rate limits and makes re-recording single paragraphs easy
• Audacity for post-processing: A 2-second “compressor” effect smooths out volume variations dramatically

What Doesn’t

• Full AI-generated podcasts (like NotebookLM) are impressive demos but lack editorial control
• Piper TTS alone sounds too robotic for long-form listening
• Free cloud TTS with limits (like Google Cloud TTS free tier) runs out fast — Edge-TTS doesn’t have a hard limit, just throttling

Where I’m Going Next

I’m working on a script that automates the entire pipeline:

• Input: Markdown script with [pause], [emphasis], [transition] markers
• Output: Full podcast episode with intro/outro music

If you’re interested, I’ll share the script on GitHub once it’s polished. Leave a comment or DM me.

Bottom Line

You don’t need to spend $50/month on ElevenLabs to have a decent AI podcast voice. The open-source ecosystem has matured to the point where free tools can produce genuinely listenable audio — you just need to know which ones to combine.

Try Edge-TTS + Audacity this weekend. You’ll have a podcast episode by Sunday.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Email *

Name Comment Message

Comment or Message

Submit