Uploaded September 2026 | Updated November 2026, 4 days ago
Build this yourself, free through the end of November, subject to the Fair Use Policy: https://fish.audio/?fpr=siraj80
This video contains a paid integration by Fish Audio.
My AI agent already does real work. It reads my inbox, drafts my replies, flags the ones that need a human and moves things on my calendar. It has been doing it for months and it talks like a log file, so I skim it, and skimming isn't checking.
So I added one layer: a voice, built on Fish Audio S2.1 Pro. One function, text in and audio out, English and Hindi in the same voice. Then I ran an experiment I didn't enjoy. One working session, approvals by voice only, log window closed. At the end I opened the log on camera and counted every decision I'd approved without reading: how many were correct and how many were wrong.
I said I'd publish the number before I knew what it was. It's 24 correct, 3 wrong.
The rule I came out with, which is yours to steal: give your agent a voice for output, never for approval. Voice is a great narrator and a terrible auditor. That's about what a confident delivery does to a listener, not about the speech model, which read every wrong sentence perfectly.
WHAT YOU'LL LEARN
How to add a voice layer to an agent you already have, in one function
Why one voice across two languages is a different thing from two voices
Why faster approval means less time to be suspicious
How to set an approval boundary that survives your own enthusiasm
CHAPTERS
0:00 I said yes without looking
0:29 What the agent already does
1:01 The missing layer
1:47 Fish Audio
2:58 It has a personality now
3:48 The rule: voice only
4:35 The Hindi thread
5:02 The calm wrong answer
5:48 Open the log
6:07 What the voice actually changed
6:36 The rule you should steal
7:43 Close
Fish Audio S2.1 Pro: https://fish.audio/?fpr=siraj80
The Developer API is free through the end of November, subject to their Fair Use Policy.
Tell me in the comments what you'd never let a voice approve.
๐ Subscribe for more real-world AI builds: youtube.com/@SirajRaval
๐ฌ Business inquiries: hello@sirajraval.com
๐ฒ Follow
X: https://x.com/sirajraval
Instagram: instagram.com/sirajraval
LinkedIn: linkedin.com/in/sirajraval
Build this yourself, free through the end of November, subject to the Fair Use Policy: https://fish.audio/?fpr=siraj80
This video contains a paid integration by Fish Audio.
My AI agent already does real work. It reads my inbox, drafts my replies, flags the ones that need a human and moves things on my calendar. It has been doing it for months and it talks like a log file, so I skim it, and skimming isn't checking.
So I added one layer: a voice, built on Fish Audio S2.1 Pro. One function, text in and audio out, English and Hindi in the same voice. Then I ran an experiment I didn't enjoy. One working session, approvals by voice only, log window closed. At the end I opened the log on camera and counted every decision I'd approved without reading: how many were correct and how many were wrong.
I said I'd publish the number before I knew what it was. It's 24 correct, 3 wrong.
The rule I came out with, which is yours to steal: give your agent a voice for output, never for approval. Voice is a great narrator and a terrible auditor. That's about what a confident delivery does to a listener, not about the speech model, which read every wrong sentence perfectly.
WHAT YOU'LL LEARN
How to add a voice layer to an agent you already have, in one function
Why one voice across two languages is a different thing from two voices
Why faster approval means less time to be suspicious
How to set an approval boundary that survives your own enthusiasm
CHAPTERS
0:00 I said yes without looking
0:29 What the agent already does
1:01 The missing layer
1:47 Fish Audio
2:58 It has a personality now
3:48 The rule: voice only
4:35 The Hindi thread
5:02 The calm wrong answer
5:48 Open the log
6:07 What the voice actually changed
6:36 The rule you should steal
7:43 Close
Fish Audio S2.1 Pro: https://fish.audio/?fpr=siraj80
The Developer API is free through the end of November, subject to their Fair Use Policy.
Tell me in the comments what you'd never let a voice approve.
๐ Subscribe for more real-world AI builds: youtube.com/@SirajRaval
๐ฌ Business inquiries: hello@sirajraval.com
๐ฒ Follow
X: https://x.com/sirajraval
Instagram: instagram.com/sirajraval
LinkedIn: linkedin.com/in/sirajraval










