WEBVTT

00:00:00.000 --> 00:00:02.640
I am, what am I called, head of digital audio

00:00:02.640 --> 00:00:07.040
at RNZ. So I run the podcast team here at New

00:00:07.040 --> 00:00:09.980
Zealand's National Broadcaster and a bunch of

00:00:09.980 --> 00:00:12.380
the weekly radio shows and our overall digital

00:00:12.380 --> 00:00:15.699
strategy. RNZ, is that kind of the same as the

00:00:15.699 --> 00:00:18.500
BBC or CBC, that sort of thing? Exactly right.

00:00:18.579 --> 00:00:23.059
We talk a lot to the BBC, ABC, CBC. We're state

00:00:23.059 --> 00:00:25.260
funded, but obviously independent at arm's length

00:00:25.260 --> 00:00:27.839
with that funding. But yes, we're a traditional

00:00:27.839 --> 00:00:30.980
public broadcaster. And you do a bunch of different

00:00:30.980 --> 00:00:33.920
podcasts, don't you? And I think you've been

00:00:33.920 --> 00:00:36.920
podcast first for quite some time. Yeah, we have.

00:00:37.000 --> 00:00:43.219
We've since about 2017. have about 140 shows

00:00:43.219 --> 00:00:46.179
on the slate probably. Half of those roughly

00:00:46.179 --> 00:00:49.280
would be repurposed radio shows, but the rest

00:00:49.280 --> 00:00:53.219
are all podcast originals or various versions

00:00:53.219 --> 00:00:56.500
thereof, some video, some not. But yes, all digital

00:00:56.500 --> 00:00:59.840
first, all very much focused on our digital audience

00:00:59.840 --> 00:01:04.719
and trying to make sure they are the glorious

00:01:04.719 --> 00:01:06.640
thing that is digital audio as opposed to just

00:01:06.640 --> 00:01:09.099
hacked up radio. Yes, which is a big difference.

00:01:11.030 --> 00:01:13.609
We'll get on to NARC in a minute, but what are

00:01:13.609 --> 00:01:16.189
some of your biggest shows that we would have

00:01:16.189 --> 00:01:18.150
heard of? Look, people might have listened to

00:01:18.150 --> 00:01:22.769
Gone Fishing and The Service, which are kind

00:01:22.769 --> 00:01:27.829
of true crime or spy shows that we have done

00:01:27.829 --> 00:01:30.769
in the past. Black Sheep is a history show that

00:01:30.769 --> 00:01:34.989
we do that has got some global renown. Yeah,

00:01:35.109 --> 00:01:37.049
that's the kind of thing we spend a lot of time

00:01:37.049 --> 00:01:39.409
on. Did Titanic Sink has been playing quite a

00:01:39.409 --> 00:01:42.500
bit in Australia. um our changing world is our

00:01:42.500 --> 00:01:44.879
science show that plays weekly on both the cbc

00:01:44.879 --> 00:01:48.540
and the abc as well as rnz so yeah we we like

00:01:48.540 --> 00:01:50.319
to think we punch them up our weight now you've

00:01:50.319 --> 00:01:55.459
just launched a new show called narc and uh it's

00:01:55.459 --> 00:01:58.359
all to do with the um the murder of a prison

00:01:58.359 --> 00:02:04.719
inmate in 1985 um i i'm interested in this show

00:02:04.719 --> 00:02:06.879
i mean firstly why don't you tell us what narc

00:02:06.879 --> 00:02:12.039
is about it's a it's a 17 episode series so we

00:02:12.039 --> 00:02:14.879
could stay here for weeks discussing it. It really

00:02:14.879 --> 00:02:17.139
is one of those true crime shows where you just

00:02:17.139 --> 00:02:20.939
think you couldn't want for anything more. It

00:02:20.939 --> 00:02:24.319
is a murder in a prison, the first in Mount Eden,

00:02:24.340 --> 00:02:27.620
New Zealand's biggest prison in 1985. Obviously

00:02:27.620 --> 00:02:29.639
the people who were witnesses or around at the

00:02:29.639 --> 00:02:32.379
time were all convicted criminals so who do you

00:02:32.379 --> 00:02:34.689
believe? The man who was ultimately convicted,

00:02:34.810 --> 00:02:37.629
not once but twice for the crime, and had his

00:02:37.629 --> 00:02:40.430
conviction quashed by the Governor -General twice

00:02:40.430 --> 00:02:44.169
for the same crime, was a guy called Rosser Pelgrin,

00:02:44.210 --> 00:02:48.090
who was a low -time burglar who was in prison

00:02:48.090 --> 00:02:50.250
at the time. The police had started talking deals

00:02:50.250 --> 00:02:53.030
and money even before he made a statement. Ultimately,

00:02:53.110 --> 00:02:55.990
it was evidence from a narc that got Rosser Pelgrin

00:02:55.990 --> 00:02:58.750
convicted of murder. Perhaps the most dangerous

00:02:58.750 --> 00:03:01.270
type of witness is the incentivised prison witness.

00:03:02.000 --> 00:03:05.379
So NARC's not only a story about the 1985 murder

00:03:05.379 --> 00:03:08.780
of Darcy Tehera, it's also the story of Rosa

00:03:08.780 --> 00:03:11.240
Pelgrin, the man twice convicted of the crime.

00:03:11.580 --> 00:03:14.500
He went to his grave in 2013, insisting he was

00:03:14.500 --> 00:03:17.479
innocent, and he did not murder Darcy Tehera

00:03:17.479 --> 00:03:20.919
on January the 6th, 1985, on a Sunday morning

00:03:20.919 --> 00:03:23.680
as they were getting breakfast. The key part

00:03:23.680 --> 00:03:27.740
in that sentence was that Rosa Pelgrin has been

00:03:27.740 --> 00:03:31.210
dead for... the best part of 13 years. But his

00:03:31.210 --> 00:03:34.310
widow is going back to court, and this is part

00:03:34.310 --> 00:03:37.050
of this news that we just broke with the release

00:03:37.050 --> 00:03:39.289
of the podcast. His widow is going back to court

00:03:39.289 --> 00:03:41.090
to try and clear his name 13 years after his

00:03:41.090 --> 00:03:43.830
death, using a precedent that was set by the

00:03:43.830 --> 00:03:46.349
Peter Ellis case here in New Zealand, which people

00:03:46.349 --> 00:03:47.870
might have heard the conviction podcast that

00:03:47.870 --> 00:03:50.189
we did about the Peter Ellis case, in which an

00:03:50.189 --> 00:03:53.370
offender was able to clear his name after he

00:03:53.370 --> 00:03:56.930
had died. This is... unprecedented territory

00:03:56.930 --> 00:03:59.069
we're getting into legally, but it's a fascinating

00:03:59.069 --> 00:04:02.969
case of a palgrin. He escaped prison at one point

00:04:02.969 --> 00:04:06.150
and went on talkback radio to insist on his innocence.

00:04:06.569 --> 00:04:12.349
He has had a remarkable story, as has the NARC,

00:04:12.469 --> 00:04:15.889
who is the man who testified to seeing a palgrin

00:04:15.889 --> 00:04:19.689
commit the murder. He has also had a remarkable

00:04:19.689 --> 00:04:24.069
life since he made that witness statement. And

00:04:24.069 --> 00:04:25.550
the question is whether we believe him or not.

00:04:25.819 --> 00:04:28.459
And you say that a pilgrim has been dead for

00:04:28.459 --> 00:04:32.199
some time, but he's on this new show, isn't he?

00:04:32.300 --> 00:04:36.120
And he's actually explaining some of the things

00:04:36.120 --> 00:04:40.120
that went on. You've clearly used AI for this.

00:04:40.459 --> 00:04:44.019
How have you used AI for this? Yes. Well, the

00:04:44.019 --> 00:04:46.959
first thing to say, I mean, to be very precise

00:04:46.959 --> 00:04:49.420
and transparent about these terms, we're careful

00:04:49.420 --> 00:04:52.240
to say he is not on it. This is AI, but it is

00:04:52.240 --> 00:04:55.120
not him. We want to be really clear that we're

00:04:55.120 --> 00:04:57.120
not trying to be cute about this. We're not trying

00:04:57.120 --> 00:05:00.680
to be in any way deceptive. And so we've spent

00:05:00.680 --> 00:05:03.319
a bit of time in episode two and then repeated

00:05:03.319 --> 00:05:05.639
some of it in later episodes explaining exactly

00:05:05.639 --> 00:05:09.600
what and why we did it. What we had in the investigation

00:05:09.600 --> 00:05:12.480
was screeds of written evidence. He had written

00:05:12.480 --> 00:05:15.879
his memoirs before he died. He had, over the

00:05:15.879 --> 00:05:18.120
years, there were transcripts from court cases.

00:05:18.240 --> 00:05:20.649
There were affidavit. signed sworn and signed

00:05:20.649 --> 00:05:24.889
there were print articles written about him so

00:05:24.889 --> 00:05:28.410
we had his voice in written form in great volumes

00:05:28.410 --> 00:05:33.509
but we had very little audio of him um some we

00:05:33.509 --> 00:05:35.449
were you could have claimed through fair dealing

00:05:35.449 --> 00:05:37.769
some we could get the rights to but just generally

00:05:37.769 --> 00:05:40.350
there was just not a lot and there was not a

00:05:40.350 --> 00:05:42.189
lot of him saying the things that he had written

00:05:42.189 --> 00:05:46.009
down so how did we meet the medium in that regard

00:05:46.009 --> 00:05:49.449
well usually we'd use an actor and That would

00:05:49.449 --> 00:05:51.829
have been common practice and was when we started

00:05:51.829 --> 00:05:53.750
working on the podcast two and a half years ago.

00:05:54.129 --> 00:05:56.129
But as we walked through, and this has been a

00:05:56.129 --> 00:06:00.269
very long investigation, as we worked through

00:06:00.269 --> 00:06:03.050
the investigation, it became clear that AI was

00:06:03.050 --> 00:06:04.550
giving us options that we didn't have before.

00:06:05.350 --> 00:06:08.129
And Mike Wesley -Smith, who's the host and lead

00:06:08.129 --> 00:06:09.829
producer on this podcast who we're working with,

00:06:10.290 --> 00:06:14.110
came to us and said, what about using AI? We've,

00:06:14.170 --> 00:06:21.519
you know, the voice. of many was, well, would

00:06:21.519 --> 00:06:23.759
his family, would his widow be okay with that?

00:06:23.819 --> 00:06:27.139
What would he think? We had those conversations.

00:06:27.259 --> 00:06:30.579
His widow, Julie, was very keen, loved the idea.

00:06:31.240 --> 00:06:34.740
We got permission from her and his estate to

00:06:34.740 --> 00:06:38.740
do that. We went through our AI working group

00:06:38.740 --> 00:06:42.399
here at RNZ to have the ethical discussions and

00:06:42.399 --> 00:06:44.680
ultimately we got to the place where we thought

00:06:44.680 --> 00:06:48.980
it was the right. thing to do and would really

00:06:48.980 --> 00:06:53.980
enhance the podcast. And so we used some of the

00:06:53.980 --> 00:06:57.699
few radio interviews that he gave and taught

00:06:57.699 --> 00:07:02.019
an AI voice cloning tool his voice. I had a fair

00:07:02.019 --> 00:07:04.220
idea of who was in on it, but I wasn't 100 %

00:07:04.220 --> 00:07:07.120
sure. There was about three or four who I thought

00:07:07.120 --> 00:07:08.860
were going to be in on it, with one or two giving

00:07:08.860 --> 00:07:10.939
him a bash, while one or two kept an eye out.

00:07:11.439 --> 00:07:13.819
When Tahira came into the kitchen, I was thinking,

00:07:13.879 --> 00:07:16.649
I don't want to be near this, just in case. you

00:07:16.649 --> 00:07:18.790
never know what's going to happen. This is the

00:07:18.790 --> 00:07:23.509
first time that AI has been used by RNZ. It is.

00:07:23.529 --> 00:07:25.889
It is. And as far as we know, the first time

00:07:25.889 --> 00:07:28.750
that we've heard in a podcast or radio show of

00:07:28.750 --> 00:07:33.050
AI being used by a dead criminal to try and clear

00:07:33.050 --> 00:07:35.470
their name. So it's a very specific use, and

00:07:35.470 --> 00:07:38.569
that was part of our consideration as well. RNZ

00:07:38.569 --> 00:07:40.949
will treat this on a case -by -case basis. But

00:07:40.949 --> 00:07:44.790
we thought... You know, if part of journalism

00:07:44.790 --> 00:07:49.490
is giving voice to the voiceless, then it's often

00:07:49.490 --> 00:07:51.709
to people who are the most vulnerable. And if

00:07:51.709 --> 00:07:54.930
you've been accused of murder, there is no greater

00:07:54.930 --> 00:07:57.870
stain on your reputation. And if you're insisting

00:07:57.870 --> 00:07:59.410
that you're innocent, and look, we don't come

00:07:59.410 --> 00:08:01.129
to conclusions. This is an investigation which

00:08:01.129 --> 00:08:03.569
raises a lot of questions about the safety of

00:08:03.569 --> 00:08:06.069
the conviction, but it certainly doesn't take

00:08:06.069 --> 00:08:08.470
sides as to whether we believe the narc or whether

00:08:08.470 --> 00:08:11.470
we believe a pilgrim. It is a balanced and fair

00:08:11.470 --> 00:08:14.180
approach to the story. We figured that a pilgrim

00:08:14.180 --> 00:08:17.579
had a right to a voice in there and to make his

00:08:17.579 --> 00:08:20.980
case, declare his name. And I guess the bottom

00:08:20.980 --> 00:08:24.540
line, James, is how much more powerful to hear

00:08:24.540 --> 00:08:28.579
a simile of his voice pleading his own case.

00:08:28.779 --> 00:08:31.459
You know, it just takes the connection with the

00:08:31.459 --> 00:08:33.500
audience to another level. You said something

00:08:33.500 --> 00:08:35.419
right at the beginning of this. Normally, we

00:08:35.419 --> 00:08:38.639
would have used an actor. And I think, you know,

00:08:38.659 --> 00:08:40.879
it's very easy to turn around and say, oh, you

00:08:40.879 --> 00:08:43.320
know, AI, this is a dreadful idea. Well, actually,

00:08:43.440 --> 00:08:47.419
this is a better idea, it seems to me, to have

00:08:47.419 --> 00:08:53.080
the real person's voice reading his own words

00:08:53.080 --> 00:08:57.440
rather than having an actor coming in and reading

00:08:57.440 --> 00:09:00.360
those words for him. I mean, how much more powerful,

00:09:00.460 --> 00:09:02.820
right? The sense that, again, you know, it is

00:09:02.820 --> 00:09:07.549
a machine ultimately. It was so moving for his

00:09:07.549 --> 00:09:09.789
family and how much more powerful as a listener

00:09:09.789 --> 00:09:15.850
to actually hear the guy's tones and, as he says,

00:09:15.850 --> 00:09:18.370
his words. I was a career criminal. People thought

00:09:18.370 --> 00:09:21.570
it was strange, but it wasn't. We still, in terms

00:09:21.570 --> 00:09:24.049
of actors, and from an audience point of view,

00:09:24.129 --> 00:09:26.070
we think it's a really good call. Obviously,

00:09:26.070 --> 00:09:29.350
there are issues around the use of actors, for

00:09:29.350 --> 00:09:33.230
example. We still use... crikey a couple of dozen

00:09:33.230 --> 00:09:35.830
actors i think there is as i say over 17 episodes

00:09:35.830 --> 00:09:38.509
there are many parts um and things to be read

00:09:38.509 --> 00:09:40.929
out so we still use actors i'm out and we still

00:09:40.929 --> 00:09:45.490
used an actor in conjunction with the um ai voice

00:09:45.490 --> 00:09:49.370
we fed the words we wanted from the memoir and

00:09:49.370 --> 00:09:53.110
the articles and so forth into the the ai tool

00:09:53.110 --> 00:09:57.389
and it very um two really interesting things

00:09:57.389 --> 00:10:02.370
one we ended up using only one minute of really

00:10:02.370 --> 00:10:06.129
high quality audio to teach the voice. There's

00:10:06.129 --> 00:10:07.429
a lot of advice out there when we started looking

00:10:07.429 --> 00:10:09.350
into it saying you'll need 30 minutes of a voice

00:10:09.350 --> 00:10:12.450
for AI to learn it. We did it with one minute

00:10:12.450 --> 00:10:17.230
of audio. We tried it with several minutes of

00:10:17.230 --> 00:10:19.610
lower quality audio and it didn't sound as good

00:10:19.610 --> 00:10:23.009
or as realistic. So we found that less audio

00:10:23.009 --> 00:10:26.149
of a higher quality is more useful to AI than

00:10:26.149 --> 00:10:30.299
more audio of a lower quality. But then when

00:10:30.299 --> 00:10:33.240
we initially fed it in, it was very, the accent

00:10:33.240 --> 00:10:36.279
was hard for AI to get and it ended up sounding

00:10:36.279 --> 00:10:38.120
either very American or very English to start

00:10:38.120 --> 00:10:43.779
with. So we used a New Zealand actor who had

00:10:43.779 --> 00:10:47.980
listened to Russell Prelgrin's interviews to

00:10:47.980 --> 00:10:52.360
try and mimic his intonation and his pace and

00:10:52.360 --> 00:10:55.779
also to get the Kiwi accent. And so it was a

00:10:55.779 --> 00:10:59.679
combination of that AI voice. but also the actors'

00:10:59.840 --> 00:11:03.279
abilities. Ross, yeah. Yeah, so you're doing,

00:11:03.399 --> 00:11:07.159
I think Eleven Labs calls it parrot mode. Which

00:11:07.159 --> 00:11:12.120
tool did you use, were you able to say? I guess

00:11:12.120 --> 00:11:14.320
not, and given you just mentioned them, we used

00:11:14.320 --> 00:11:18.320
Eleven Labs. All right, okay. Well, that's nice

00:11:18.320 --> 00:11:21.279
and easy. I'm guessing this isn't in video. This

00:11:21.279 --> 00:11:24.419
is just purely an audio show? Purely audio. Yep,

00:11:24.460 --> 00:11:28.879
purely audio. It's 17 episodes. And the episode

00:11:28.879 --> 00:11:33.120
is, I don't know, 35 to 65 minutes each. So there

00:11:33.120 --> 00:11:34.940
is a lot of audio. And this is the point, right,

00:11:35.019 --> 00:11:38.179
that it would have been, there's a lot of Ross

00:11:38.179 --> 00:11:40.879
Appelgren in there, a lot of him trying to tell

00:11:40.879 --> 00:11:44.600
his own story. And an actor is great, but...

00:11:44.940 --> 00:11:48.700
over that kind of duration. And when you're trying

00:11:48.700 --> 00:11:53.460
to, you know, a man's trying to plead for, not

00:11:53.460 --> 00:11:54.820
for his life, we don't have the descendants in

00:11:54.820 --> 00:11:59.379
New Zealand, but plead his innocence. It's just

00:11:59.379 --> 00:12:02.220
that much more of a connection, I think, for

00:12:02.220 --> 00:12:04.259
the audience to hear him speak himself. Really

00:12:04.259 --> 00:12:07.139
interesting to hear about AI being used in this

00:12:07.139 --> 00:12:10.419
way. NARC is available wherever you get your

00:12:10.419 --> 00:12:13.899
podcasts, I'm assuming? It sure is. And it's

00:12:13.899 --> 00:12:15.519
rolling out between now and the first week of

00:12:15.519 --> 00:12:18.679
December, three episodes a week. And I tell you,

00:12:18.720 --> 00:12:21.100
it's the best true crime story I've heard for

00:12:21.100 --> 00:12:23.940
a very long time. Tim, thank you so much. Thank

00:12:23.940 --> 00:12:24.080
you.