Latest / Business Podcasting: Production Strategy / How do you connect AI to your podcast tools?

Transcript

How do you connect AI to your podcast tools? · Business Podcasting: Production Strategy

• 0:00
  [Synthetic version of Lam Hoang's voice]
• 0:07
  This is Business Podcasting: Production Strategy, produced by WorkHacker.
• 0:13
  I'm your host, Lam Hoang, co-founder of WorkHacker.
• 0:16
  You're hearing a synthetic version of my voice.
• 0:19
  Each episode I take one question about making a podcast for your business,
• 0:23
  tell you what I would do and how solid the evidence is.
• 0:32
  Today's question: how do you connect AI to your podcast tools?
• 0:37
  Connect one repeated job through the route your existing tools support.
• 0:41
  As of October twenty twenty-six, voice providers say their
• 0:44
  connections can generate speech and return a download link.
• 0:48
  This lets the assistant return a recording for you to hear and revise.
• 0:52
  Here's how I use that loop.
• 0:54
  An approved segment goes to the voice service and the audio comes back.
• 0:59
  I listen against the script.
• 1:00
  Say a name is pronounced wrong.
• 1:02
  Tell the agent the spelling is right; the pronunciation is wrong.
• 1:06
  The agent proposes a pronunciation fix instead of rewriting the sentence.
• 1:11
  After approval, a new draft comes back to hear in context.
• 1:15
  Then the whole episode gets a listen before release.
• 1:18
  Here's what you'll leave with.
• 1:20
  You'll know how to get a draft back from your tools and
• 1:22
  improve it with the agent before anything goes out.
• 1:26
  First: what job are you giving the tool?
• 1:29
  In that example, it is turning approved words into a recording you can play.
• 1:34
  The assistant needs to send the text and return the result,
• 1:37
  not just explain how to open the voice service.
• 1:40
  Choose a handoff you already repeat,
• 1:42
  so the connection takes a recognizable job off your list.
• 1:45
  The connection is useful because you can keep working
• 1:48
  on the same draft when the result comes back.
• 1:51
  The script is still beside you and the agent still has your instructions.
• 1:55
  You can point to the passage that needs work instead of starting
• 1:58
  another conversation with the whole episode pasted in again.
• 2:02
  The agent can help prepare the correction, but receiving
• 2:04
  an audio file does not mean it has heard the recording.
• 2:08
  That is why the feedback in the name example matters.
• 2:11
  Asking the agent to improve the passage leaves it to
• 2:14
  guess whether you mean the words or their delivery.
• 2:18
  Saying the spelling is correct gives it a clear boundary for the fix.
• 2:22
  Ask it to show what it wants to change,
• 2:24
  then hear the result with the sentences on either side.
• 2:28
  A corrected name should still sit naturally in the episode.
• 2:32
  Check whether the tool can make the correction you want.
• 2:35
  As of October twenty twenty-six, Cartesia documents speech
• 2:39
  generation and pronunciation dictionaries through its MCP tools.
• 2:43
  A connection that generates audio may therefore
• 2:46
  also offer a way to specify how a name is spoken.
• 2:49
  If your connection lacks that action, the agent can
• 2:52
  prepare a correction for you to apply in the service.
• 2:55
  You can still use the same review loop while handling that step yourself.
• 2:59
  The useful connection is the one that gets your actual job done.
• 3:03
  For a scripted segment, ask whether it can return audio from
• 3:07
  approved text and support the changes you make after listening.
• 3:11
  If your show is interviews, bringing back a
• 3:13
  transcript may be the handoff you repeat more often.
• 3:16
  Choose the job before choosing the connection.
• 3:19
  Second: which connection fits your tools?
• 3:23
  Look at what your assistant supports and what your
• 3:25
  existing provider offers for the job you chose.
• 3:29
  A built-in app may cover it already; an official MCP
• 3:32
  connection or an API integration may be the next route.
• 3:36
  An API is a service's way of taking requests from software.
• 3:40
  MCP means Model Context Protocol and lets
• 3:44
  assistants find and call tools in a common way.
• 3:47
  A connector packages that access inside the assistant you use.
• 3:51
  The useful distinction is whether a connection can act
• 3:54
  on your file or only help you read about the service.
• 3:58
  For instance, AssemblyAI documents a server that
• 4:01
  helps an assistant search its documentation.
• 4:04
  That server explains transcription; it does not transcribe your recording.
• 4:09
  Processing the file needs the separate action route or API.
• 4:13
  For the voice loop, look for the action that sends text and returns audio.
• 4:18
  For voice work, there are several options to investigate.
• 4:21
  As of October twenty twenty-six, ElevenLabs and Cartesia
• 4:25
  document official MCP routes for generating speech.
• 4:29
  Resemble AI and Google Cloud document speech generation through APIs.
• 4:34
  Choose a voice that fits your show and confirm
• 4:36
  the connection your assistant can use.
• 4:38
  Use current instructions because the supported route can change.
• 4:42
  ElevenLabs marks its older local MCP server as
• 4:46
  deprecated and points to the hosted connection.
• 4:49
  That notice was checked in October twenty twenty-six.
• 4:52
  Following an old tutorial can therefore send you
• 4:54
  toward a connection the provider has replaced.
• 4:57
  If your provider offers an API without a suitable connector,
• 5:01
  have an integration built for that specific operation.
• 5:04
  Give the person building it the input,
• 5:06
  the result and the actions that should require your approval.
• 5:10
  For a speech request, the API documentation
• 5:13
  describes how text goes in and audio comes back.
• 5:16
  Start with that defined handoff before connecting another production step.
• 5:21
  The connection you choose should match the draft you want back.
• 5:24
  A product name or a list of available tools is only the starting point.
• 5:29
  Check the operation and what it returns before you authorize the request.
• 5:33
  Third: what can the connection do on your account?
• 5:37
  Asking for one voice draft is a small task,
• 5:40
  but the account grant may allow much more.
• 5:43
  Inspect the requested permission and available actions before authorizing it.
• 5:48
  You want to know whether that connection can change
• 5:50
  other work as well as return your recording.
• 5:53
  An instruction in chat does not change the
• 5:55
  permissions you granted to the connection.
• 5:57
  Naming one folder does not necessarily restrict access to that folder,
• 6:02
  and asking for reading alone does not remove write permission.
• 6:06
  Where your account has tool controls, disable actions the job
• 6:09
  does not need and retain confirmation for consequential changes.
• 6:14
  The way a service returns a file can make that permission decision concrete.
• 6:18
  As of October twenty twenty-six,
• 6:20
  Descript says its API supports imports and editing jobs.
• 6:24
  Its help page says a rendered audio download
• 6:27
  requires publishing a Descript web link.
• 6:29
  That makes obtaining the finished file a sharing decision too.
• 6:33
  If you need a private export, use the app's export route
• 6:36
  instead of authorizing publication just to get the download.
• 6:40
  The cost belongs in that loop too.
• 6:42
  Generating the first recording uses the service,
• 6:45
  and generating the corrected version uses it again.
• 6:48
  Check what each generation costs on your account before asking for another draft.
• 6:53
  As of October twenty twenty-six, Descript warns that API
• 6:57
  and MCP work can consume media minutes and AI credits.
• 7:01
  An instruction to keep improving the episode
• 7:03
  is not an agreed budget for repeated requests.
• 7:07
  Treat an API key like a password and put it in the connection's supported settings.
• 7:12
  A page or transcript can contain instructions designed to redirect the agent.
• 7:16
  Treat those words as source material, with your actual
• 7:19
  task and account controls deciding what happens next.
• 7:23
  Last: how do you try the loop?
• 7:26
  Start with a short approved segment if generating
• 7:29
  your own voice is the job you want to connect.
• 7:32
  Listen to what comes back, give the agent one specific correction
• 7:36
  and review the proposed change before another request.
• 7:39
  That is enough to find out whether the connection
• 7:41
  handles the handoff you meant to remove.
• 7:44
  Not using a synthetic voice?
• 7:46
  Try the same idea with an existing transcript and a draft of show notes.
• 7:51
  As of October twenty twenty-six, the assistant vendors
• 7:54
  document storage connections for supported accounts.
• 7:57
  Choose available storage you already use and inspect the grant before connecting.
• 8:02
  Keep confirmations enabled and block unnecessary actions where your account permits.
• 8:07
  Ask for notes from one selected transcript,
• 8:10
  with references back to the passages used.
• 8:13
  Keep the draft in the conversation and review names,
• 8:16
  quotations and meaning against the original.
• 8:19
  If the access is too broad or the connector is unavailable,
• 8:22
  upload the transcript manually for this trial.
• 8:25
  You can test the drafting and correction while
• 8:28
  looking for a connection whose grant fits.
• 8:30
  Whichever draft you try, count the review
• 8:33
  time alongside the copying step you removed.
• 8:36
  Did the agent bring back the right result, understand the
• 8:39
  correction and preserve the parts you had already approved?
• 8:42
  Those are things you can check on your own episode,
• 8:45
  without assuming that a successful tool request means successful production.
• 8:49
  Running voices on your own machine is another route,
• 8:52
  with software and voice models to choose and maintain.
• 8:55
  If you'd like an episode on running voices on your own machine, tell us.
• 9:00
  So choose one job and a connection that can return the draft you need.
• 9:04
  For the voice loop, send approved words and listen
• 9:07
  to the recording before deciding what to change.
• 9:10
  Let the agent help with that correction, then hear the
• 9:13
  new passage in context and check the finished episode.
• 9:17
  The tools carry the work between those decisions;
• 9:20
  you decide what is ready to go out.
• 9:22
  A free prompt for your own AI to check which of your podcast
• 9:25
  tools it can connect to comes with the written version.
• 9:28
  And that's it for today's question.
• 9:30
  Thanks for listening.
• 9:31
  The written version is on workhacker.com, with every source.
• 9:35
  And if you want a podcast like this for your business, WorkHacker makes them.
• 9:39
  See you next time.