Fish Audio Drama 3: direct AI voices like an actor 🎭
Three builds: an AI receptionist, a daily podcast in your voice, voiceovers that act.
Hey, it's Adam 👋 You commented DRAMA on my Fish Audio reel. Here's the link to try Drama 3, plus the step-by-step for all three builds from the video: an AI receptionist, a daily podcast in your own voice, and voiceovers that actually act.
📬 Get guides like this in your inbox every Thursday
Read it here free, or subscribe and get every new one, plus the AI news that matters.
Try Fish Audio Drama 3
In partnership with Fish Audio · #FishAudioPartner #ad
3 things to build with it this week 🎭
Drama 3 is Fish Audio's new voice model. Instead of hoping the AI reads your line the right way, you tell it how — in brackets, in plain English, the way you'd direct a voice actor. It performs it.
- 1An AI receptionist that changes tone per caller.
- 2Your own daily podcast, in your own voice.
- 3Voiceovers that actually act.
Drama 3 is in early access for a limited time.
How directing works (read this first) 🎬
Everything in this guide uses one idea: tags in brackets.
[calm, reassuring] I completely understand. Let me sortthat out for you right now.
What you can do with tags:
- Write the emotion in plain English. Not just
[happy]— try[barely holding it together, forcing a smile]. The more specific, the better the performance. - Change delivery mid-sentence. A tag affects everything after it, so where you put it matters.
- Write whole scenes. Multiple speakers in one script, generated in one pass.
- Fix one word, not the whole clip. If a single word comes out wrong, regenerate just that word.
- 83+ languages. Same tags, any language.
- Your voice or theirs. Clone your own voice, or pick one from the Fish Audio library.
Position matters — try this:
I got the results back. [pause] [quiet, stunned] Weactually did it. [excited, building] We actually did it!
Same words, three different moods, because the tags sit in different places.
Build 1 · An AI receptionist that changes tone per caller ☎️
Claude builds the agent, Fish Audio gives it the voice. The tags are what make it sound human: calm with the angry caller, warm with the nervous one.
What you need
- Claude (desktop app, paid plan, Code tab or Cowork)
- A Fish Audio account and API key → fish.audio
- Everything about the business: services, prices, hours, location, FAQs, booking rules
Step 1 · Write the business brief
Put it all in one document. The more Claude knows, the fewer things the receptionist makes up.
Business: [NAME]What we do: [SERVICES + PRICES]Hours: [HOURS]Location: [ADDRESS + PARKING]Booking rules: [HOW BOOKINGS WORK, CANCELLATION POLICY]FAQs: [TOP 10 QUESTIONS + ANSWERS]Never do: [e.g. give medical advice, promise refunds,quote prices not on this list]Hand off to a human when: [e.g. complaints, emergencies,anything not covered here]
Step 2 · Store your Fish Audio key safely
Create your API key in your Fish Audio account. Don't paste it into the chat. Ask Claude to create a .env file, then open it yourself and paste the key in.
Step 3 · Let Claude build it
Attach your brief and paste:
Build me a browser-based AI receptionist for the businessin the attached brief.- It answers spoken questions, only uses facts from the brief, and says "let me get someone for you" when it doesn't know.- It speaks using Fish Audio's Drama 3 model. Read the Fish Audio docs to set it up. My API key is in .env — don't print it.- Before every reply, it detects the caller's mood and adds Fish Audio emotion tags in brackets to its own reply: angry caller → [calm, slow, reassuring] nervous caller → [warm, gentle, patient] happy caller → [friendly, upbeat] confused caller → [clear, slower, patient]- Use the voice I choose: [VOICE NAME FROM FISH AUDIO LIBRARY].Show me the plan before you build, then give me a page Ican open and talk to.
Step 4 · Test it like a real customer
Call it angry, call it nervous, ask something that isn't in the brief.
Say "I've been waiting 40 minutes and nobody called me back." It should slow down, stay calm, and not make promises the brief doesn't allow.
This is a demo you run in the browser. Putting it on a real phone line is a separate build (you'd need a telephony service). Test it thoroughly before any real customer hears it.
Build 2 · Your own daily podcast, in your own voice 🎧
Every morning, Claude reads your inbox and notes, writes a short brief, and Fish Audio reads it to you as you. I listen to mine on the way to the gym.
What you need
- Claude with Cowork (desktop app, paid plan)
- Your email connected to Claude (e.g. the Gmail connector) and your notes in a folder or connected app
- A Fish Audio account and API key → fish.audio
- A short, clean recording of your own voice
Step 1 · Clone your voice
In Fish Audio, create a voice clone from a short sample of you talking. Record somewhere quiet, speak naturally, no background music.
✋ Only clone your own voice, or a voice you have clear permission to use.
Step 2 · Connect your inbox and notes
In Claude, connect your email and give Cowork access to the folder where you keep notes (or connect your notes app).
Step 3 · Build the daily brief
Every morning, make me a 3-minute audio briefing:1. Read my emails from the last 24 hours and my notes in [FOLDER/APP].2. Write a short script: the 3 most important things today, anything I need to reply to, and one reminder from my notes.3. Write it the way I talk, casual and direct. Add Fish Audio Drama 3 tags so it sounds natural: [upbeat] for the opener, [serious] for anything urgent, [relaxed] for the wrap-up, [pause] between sections.4. Generate the audio with Fish Audio using my cloned voice [VOICE NAME]. My API key is in .env, don't print it.5. Save it as an MP3 named with today's date in [OUTPUT FOLDER].Don't send, reply to or delete any emails. Only read them.Run it once now so I can hear it.
Step 4 · Make it automatic
Once you like the first one:
Turn this into a scheduled task that runs every day at[TIME].
Listen to the first brief. If it sounds too formal, tell Claude "make the script sound more like me talking to a friend" and save that into the task.
Build 3 · Voiceovers that actually act 🎙️
Same script, completely different video. Tag it, and your faceless channel stops sounding like a robot reading a Wikipedia page.
Step 1 · Write your script normally
Don't think about tags yet. Just get the words right.
Step 2 · Let Claude direct it
You're a voice director. Add Fish Audio Drama 3 emotiontags in brackets to this script so it sounds like a realnarrator, not a robot.- Use specific, plain-English directions, not just [happy] or [sad].- Put tags exactly where the delivery should change. Tags affect everything after them.- Add [pause] before reveals and punchlines.- Hook: grab attention in the first line. Ending: land it, don't trail off.- Don't change any of my words.Script:[PASTE SCRIPT]
Step 3 · Generate in Fish Audio
Paste the tagged script into Fish Audio, pick Drama 3, choose your voice (your clone or one from the library) and generate.
If one word comes out wrong, regenerate just that word instead of the whole clip.
Generate the same script twice — once with no tags, once tagged. Play them back to back. That's the difference your viewers will hear.
Tags worth stealing 🏷️
[excited, talking fast][deadpan][whispering, like sharing a secret][pause][confident, slower, every word lands][laughing slightly][serious, low voice][surprised, then amused]
Rules I'd follow 🔑
- Be specific.
[tired but trying to sound positive]beats[sad]every time. - Tags affect what comes after them. Move a tag and you change the line.
- Keep your API key in
.env. Never paste it into a chat. - Your voice, your permission. Only clone voices you own or have the right to use.
- Receptionist = demo first. Test every scenario before a real customer hears it.
Try Fish Audio Drama 3
In partnership with Fish Audio · #FishAudioPartner #ad
Links 🔗
- Try Fish Audio Drama 3— fish.audio — partner link
- 🎬 Watch the reel— Instagram — @adam.godigital
Made in partnership with Fish Audio · @fishaudio_ai · #FishAudioPartner
Made by @adam.godigital. Follow for more AI workflows.
Liked this? Get the next one in your inbox.
Every Thursday I research the latest AI systems and workflows, then break the best one down into simple steps, with prompts you can copy. Plus the AI news that changes what you can do.