バイリンガルニュースをAmazon TranscribeとIBM Watsonに通して文字起こしの実用度を検証してみた

ポッドキャスト界において長年にわたってトップを走り続けるMami & Michaelのバイリンガルニュース。二人のニュースに対するバランス感覚と不思議なセンスでピックアップされる科学技術論文が、英語学習者にとどまらずに広くリスナーを集める魅了なんだと思います。

スポンサーをつけずに運営しているこの番組の収益は、スマホとMacのアプリになっている文字起こしアプリの月額240円のサブスクリプションから来ています。

各話の文字起こしは外注しているとのことですが、逐語(verbatim)テープ起こしの相場は1分3ドルが相場と言われているので、現在平均して90分ある1話あたりの文字起こし作成コストは3万円を超えているはずです。しかも1話ごとの買取ではなく過去分も読み放題になっているので、サブスクライバーを純増で伸ばし続けていかなければペイしないかなり厳しい課金モデルになっています。

1リスナーとして何か提案できないものかと、先日発表され、現在パブリックプレビュー中のAmazon Transcribeでこのコストを圧縮できるのかどうか、その実用度を検証してみました。

すでに英語のテープ起こし業者はニュアンスコミュニケーションズのDragon Speechなどの音声認識ソフトウェアを使用してdraftingを行なっているところが多いですが、結論から言うと、Amazon Transcribeは職業トランスクライバーの技術的失業を引き起こす段階まで進んでいるサービスです。

Amazon Transcribeを使ってみる

プレビュー段階での機能と性能であることをお断りしておきますが、以下の通り使ってみました。

AWS re:Inventでのプレゼンをみるとよくわかりますが、Transcribeは単なるSpeech to Textのサービスではなく、AWSのマネージドサービスを組み合わせてAPIドリブンなアプリケーションを作成できることを前提にしています。

AWS Step Functionsを使ったオートメーションの例

コールセンターの先行事例では、文字起こし、テキストマイニングからベストプラクティスライブラリの生成までマネージドサービスで構成しています。このように、Transcribeは自然会話の音声認識、意味認識の双方からの要求に耐えられる全体最適の精度が備わっています。

S3に適切な権限設定を施して、音声ファイルをアップロード、その保存先を指定してジョブを作成します。

ほどなく処理が完了し、結果はJSONファイルに吐き出されます。transcriptのあとに、各単語ごとの発話タイムスタンプと精度分析が配列されています。シンプルですが、SDKでタイムスタンプを利用した加工やComprehendQuicksiteへのデータ送信に適したフォーマットです。

その実力を定量、定性的に検証してみる

バイリンガルニュースで音声とテキストが一般公開されている第3回(2013.05.29)のエピソードを題材にします。Michaelのニュース読みがたどたどしかったり、Mamiがクラブ通いしてたりと、なかなか隔世の感があります。

比較対象としては、IBM WatsonのSpeech to Textを使用しました。まずは料金から。

Amazon Transcribe IBM Watson
USD 0.0004/sec

(USD 0.024/min)

USD 0.02/min

(カスタム語彙を使用する場合はUSD 0.05/min )

Transcribeで予定されているカスタム語彙モデルがサポートされた場合に追加の料金が必要になるかでコストに差が出て来ますが、現状ではそれほど変わりません。

それぞれのテキストから、Michaelの発話部分を抜き出して揃えます。

オリジナル Amazon Transcribe IBM Watson
So on May 26th at 2am, three club executives at the club Vanity in Roppongi were arrested. About 50 cops showed up and arrested them. And they were charged with violating The Entertainment Business Law. And Entertainment Business Law is that basically you can’t dance after 1am, but they had registered as a restaurant, so they were kind of exploiting a loophole, but the cops decided that they were gonna... they weren’t gonna let that fly so they showed up arrested these executives. They were going ‘til 5am, right?
Yeah, it’s really strange that they have this signs in there saying, no dancing, but you know everyone is dancing and it’s kind of what do you... what constitutes as dancing? Like, is bobbing your head dancing?
So do the cops show up and see if people were dancing first or did they just assume that people were dancing?
I wonder why Vanity and not some of these other establishments?
I mean you can go to karaoke or you know, there’s other things to do.
Yeah, you’re clearly from Tokyo. I don’t, you know this seems like they don’t have that law outside of Tokyo, do they?
So Yahoo auctions ended up sending the info of, the private information of a thousand four hundred and twenty- seven people to eight hundred and thirty- five people. And we don’t really know how this happened, right?
Yeah, like how does that happen? How does one individual have access to all that stuff? Was it just e-mails or account information?
That stuff is really, it’s really tricky, though. I’m surprised that there aren’t more mistakes in IT like this. It’s so complicated. You know, it’s so complicated, there’s so many things going on and so dynamic, they are always having to update their systems and there are so many security issues and there’s so many like really clever hackers out there. And it’s just a, it’s such a zoo.
Yeah, that dude was sweating bullets.
So basically these cockroaches are evolving. And the way that they are evolving is we have these little pesticide pellets that have a little glucose covering on them, which is what entices the cockroaches to eat these pesticide pellets and die. And these cockroaches have, they are beginning to no longer eat the pellets and we’re calling it evolution. But you know what’s really happening is that all the cockroaches that had a sweet tooth are, they’re dead, you know.
So as the ones that didn’t have that sweet tooth for that glucose covering on the pesticides as they survive. You know, they’re the ones that are repopulating, so it has kind of shifted that way. It’s kind of similar the way that viruses, not viruses, but bacteria develop a resistance to antibiotics. You know if you kill off all the ones that it’s effective against, but it’s not effective against all of them, then all of the strong ones survive and you know, that in a sense is evolution.
It’s really interesting because we tend to think that evolution is an entirely natural process and that doesn’t continue... it’s not really going on anymore. You know because of the way that we’ve, because of human intervention. But in a way we’re actually causing the things around us to adopt, possibly in a more rapid way than they would have otherwise. I imagine this is going to happen to not just that, that particular type of cockroach, but I imagine it’s happening with all the varieties of cockroach that we use pesticides against.
South Korea shut down two more nuclear reactors over fake certificates. This is something that also happened last November. They shut down two reactors then. Basically these certificates for, they’re certificates kind of insuring the parts of the reactors are up to par that their quality is high enough to withstand stresses of being part of the nuclear reactor. And the certificates were fake, so they are having to go through and get the proper parts, in this case it’s cables. So they are going to experience electricity shortages and rolling blackouts like never before and the nuclear reactors are about the third of South Korea’s power mix, and they’re gonna be closed for about four months.
Yeah, I think it’s, I think this just goes to show that this is what people do, I mean this is what humans do. We’re always gonna cut corners on these sorts of things. You know it was a very similar situation with TEPCO. And they’ve had similar problems in the United States.
Yeah, they have the sort of issue with earthquake proofing, too. Earthquake standards.
Yeah. So, I mean it’s really not what you want to be cutting corners.
So, some 600 year-old plants are basically coming back to life. These are ancient plants. They were trapped underneath glaciers, and as with the climate change, a lot of these glaciers are moving and revealing things that have been covered for the last 600 years. These are bryophytes, the plants. Bryophytes are like little ferny like not trees or anything, like smaller plants. And they were previously assumed dead. Scientists had assumed that all the plants that were underneath this, that were in this sub glacier region were dead. But as the glacier moved, they realized that these plants are actually still alive and they are actually reanimating. And these glaciers have been there since the little ice age which was 1550A.D.
Yeah, it’s nuts, they’re definitely gonna be studying these plants to see if there’s anything weird about them. Because we don’t have anything, I don’t think we have anything that old that is alive.
I wonder, I mean they wouldn’t still be alive if the glaciers weren’t there, right?
On May 31st, it’s World No Tobacco Day.
This was created in 1987 to promote awareness of the I guess not just the deaths, but all the, obviously tobacco is not good for you, so an awareness of all of those negative effects maybe try to inspire some people to quit smoking. As of the latest studies, it’s estimated that tobacco causes 5.4 million deaths every year. That’s a pretty big number, it’s bigger than pretty much every other number out there.
Yeah. 5.4 million deaths.
Well, that’s ‘cos you’re eating, It’s not because you’re not smoking. I mean supposedly there are some cognitive benefits to nicotine, but the cost far outweighs the benefits. It’s pretty clear at this point.
I mean if they came out with tobacco today, like it would never be approved. There’s no way.
There’s no way! And you know, there are so many other substances and drugs that are illegal but none of them are as bad as tobacco. None of them as harmful or none of them kill as many people, none of them are as addictive. They say that tobacco is more addictive than you know even the hardest drugs out there.
Yeah, it’s pretty mind blowing. So don’t smoke on May 31st.
so i made twenty six two am three club executives the club vanity and bundy were arrested about that did. Cops showed up and arrested them and they were charged with violating the entertainment business law and the entertainment business law is that basically you can't dance after one a m but they had registered as a restaurant so they were kind of exploiting a loophole the cops decided that they were gonna they weren't gonna let that lie so they they showed up and arrested. These is a executives they were. They were going till five a.
yes it's really strange that they they have this science and they're saying no dancing. But you know everyone's dancing you know and it's kind of what do you what constitutes is dancing like is bobbing your head dancing?
the cops show up and see if people were dancing first or did it just assume that people were dancing ?
i wonder why vanity and some of these other establishments
i mean you go cut ok that others are munitions in there.
You're clearly from tokyo i don't you know this it seems like they won't have that law outside do they ?
So yahoo auctions and it up sending the import of private information of a thousand four hundred twenty seven people two eight hundred thirty five people and we don't really know how this happened.
like you how does that ? How does how does one individual have access to all that stuff ? Was it just the e mails or like account information ?
that stuff is really it's really tricky. I'm surprised that there are more mistakes in i. T like this that's it's so complicated. So complicated there's so many things going on. It's so dynamic they're always having to update their systems and they're so many security issues and they're so many alike. Really clever hackers out there and it's just it's such a zoo.
yeah they do those his press what bullets
So basically these cockroaches are involving and the way that they're evolving ins we have these little has decide pellets that air that have little glucose covering on them. Which is what entices the cockroaches to eat. These has aside pallets and die and these these cockroaches have they're beginning to no longer eat the palace and we're calling it evolution. But you know what's really happening is that all the cockroaches that had a sweet to there ? They're dead. You know.
So as that the ones that didn't have that sweet tooth for that glucose covering on the pesticides as they survived you know they're the ones that repopulating so it is kind of shifted that way. It's kind of similar in a way that virus not viruses but bacteria develop resistance to antibiotics.
You know if you kill off all the ones that it's effective against but it's not affected. Begins all with them that only the strong ones survive.
And you know that in a sense is evolution
commodity is really interesting because we took we tend to think that evolution isn't entirely natural process and that it doesn't continue. It's not really going on anymore. You know because if because of the way that we because of human intervention but in a way we're actually causing the things around us to adapt possibly in a in a more rapid way then it would then they would have otherwise. And i i imagine this is going to happen not just that that particular type of coverage but i imagine it's it's happening with with all the varieties of cockroach that we use pesticides khomeini
south korea shut down two more nuclear reactors over fake certificates this is something that also happened last november. They shut down two reactors that basically these certificates for their certificates kind of ensuring that the parts of the reactors are are up to par that their quality is high enough. To withstand the stresses of being part of the nuclear reactor and the certificates are fake so they're having to go through it and get the proper parts in this case is it's cables so they're going to experience electricity. Electricity shortages and rolling blackouts um like never before and the nuclear reactors are about a third of south korea's power mix and they're gonna be closed for about four months
Yeah i think it's i think this just goes to show that this is what this is what people do. I mean this is what humans do and i always get that. Cut corners on these things because there were you know it was a very similar situation with that guy. You know they had similar problems in the united states.
yeah they they have they have the sort of issue with earthquake proofing to leg standards
so it's really it's really where you want to be cutting corners
so some six hundred year old plants are basically coming back to life.
These are ancient plants they were trapped underneath glaciers and as these with with climate change a lot of these glaciers are are moving and revealing things that have been cover for the last six hundred years.
These air bridal fights the plans brian fights are like little army like that. Not not trees anything like smaller plants and they were previously assumed dead. Scientists had assumed that all the plants that were underneath this that were in this sub glacier region were dead.
But as the as the glacier move they realized that these plans were actually still alive and they're actually re animating and these glaciers have been there since the little ice age which was fifteen fifty eighty.
yeah it's definitely going to be studying these plans to see if there's anything weird about because we don't have anything i don't.
I don't think we have anything bad old.
i wonder I mean they wouldn't still be alive if the glaciers weren't there.
on may thirty first it's world no tobacco day.
this was created in nineteen eighty seven to promote awareness awareness of the i guess not just the deaths but you know all the obviously tobacco's not very good for you so i know where so all of those negative effects and may be tried. Inspire some people to quit smoking so as of the latest studies it's estimated that about the cause is five point four million deaths every year so that's a pretty big number it's bigger than pretty much every five point five million five point four million deaths her years ago.
But the sundry to meet you think i can do.
Well that's cause you're eating. It's not because you're it's not because you're smoking. I mean they're supposedly there. Some cognitive benefits to nicotine. But the costs far outweighs the benefits. it's pretty clear at this point
i mean if they if they came out with tobacco today but it was never be approved. It would not be here.
no way other so many other substances in drugs that are in the open none of them are as bad as tobacco. None of them are is harmful or none of them.
Kill as many people not of a morris addictive they say that tobacco is more addictive than you know. Even the hardest drugs out there.
it's pretty mind blowing so don't smoke. On may thirty first
view raw transctibe.txt hosted with ❤ by GitHub
Onto a. Sun may twenty six at two AM. Three club executives at. The the club vanity and open the gate were arrested. About fifty cops showed up.
And arrested them and they were charged with violating the entertainment business law. And the entertainment business laws that basically you can't dance after one AM but they had registered as a restaurant so they were kind of exploiting a loophole. But the cops decided that they were gonna. They were gonna let that fly so they they showed up and arrested these exact executives. They were they were going to five AM right.
Yeah it's it's really strange that they they have the signs and they're saying no dancing. Button you know everyone's dancing you know and and it's kind of. What do you what constitutes is dancing like is bobbing your head dancing is. Could have been a.
So did the cops show up and see if people were dancing first or did they just assume that people were dancing.
I wonder why vanity and not some of these other establishments.
I mean you can go to cuddle care. You know there's a.
Yeah year. Clearly from Tokyo. I don't you know this. Seems like they want they don't have that law outside of Tokyo to the. The.
So Yahoo auctions. And I'm sending the info of. The private information.
Of a thousand four hundred twenty seven people. Eight hundred thirty five people. And. We don't really know how this happened.
Like you how does that happen. How does how does one individual have access to all that stuff. Was it just the emails or like account information.
Israeli. It's really tricky though I'm surprised there aren't more mistakes in the IT like this. Now it's so complicated there's so many things going on it so dynamic they're always having to update their systems and. There are so many security issues and there's so many like. Really clever hackers out there. And. It's just that such as soon.
Yeah. Yeah that was. As far as what bullets.
So. Basically these cockroaches are involving. And the way that they're involving is we have these little. Pesticide pellets. Better to have a.
A little glucose covering on them which is what entices the cockroaches to eat these as a side pallets and die. And these. These cockroaches have they're beginning to no longer eat the palace. And. We're calling it evolution but you know what's really happening is that all the cockroaches that had a sweet tooth there their debt. You know
so as the. The ones that didn't have that sweet tooth for that glucose covering on the pesticides as they survive. You know they're the ones that repopulating so. And it's kind of shifted that way it's kind of similar to the way that. Virus not viruses but bacteria. Develop resistance to antibiotics you know if you kill off all the ones that it's effective against but it's not effective against all of them then only the strong ones survive and I know that in a sense is evolution.
It's really interesting because we we tend to think that evolution. Is an entirely natural process and then it doesn't continue it's not really going on anymore. You know because it's because of the way that we. Because of human intervention. But in a way where actually causing the things around us to adapt. Possibly in a in a more. Rapid way than it was than they would have otherwise. And I I imagine this is going to happen to. Not just that. That particular type of cockroach but. I imagine it's it's happening with. With all the varieties of cockroach that we. Use pesticides again.
South Korea shut down two more nuclear reactors over fake certificates.
This is something that also happened last November they shut down two reactors that. Basically. Certificates for their certificates kind of.
Ensuring that the parts of the reactors are up to par that their quality is high enough. To withstand the stresses of being part of a nuclear reactor. And the certificates are fakes so they're having to go through and and get the the proper parts in this case cables. So they're going to experience electricity electricity shortages and rolling blackouts. Like never before. And. The nuclear reactors are about a third of south Koreans power mix. And they're gonna be close for about four months.
Yeah I I think it's. I think this just goes to show that this is what.
This is what people do I mean this is what humans do I always gonna.
Cut corners on these sorts of things as they were you know it was a very similar situation with KEPCO. And they had similar problems in the United States.
Yeah. They have they have they have the sort of issue with earthquake proofing to standards.
Mmhm. So I mean it's it's really it's really not where you want to be cutting corners.
Yeah. So some six hundred a year old plants. Are basically coming back to life. These are ancient plants they were trapped underneath glaciers.
And as these. With with climate change a lot of these glaciers are are moving and revealing things that have been covered the at the last six hundred years these are bryophytes the plants. I'm bryophytes are like little. Bernie. Like. Not not trees or anything like smaller plants.
And they were previously assumed debt scientists had assumed that all the plants that were underneath this. That were in this subglacial her region where dad but as the as the glacier moved they realize that. These plants were actually still alive and they're actually re animating. And. These questions have been there since the little ice age. Which was fifteen fifty eighty.
Yeah. It's not that definitely gonna be starting his plans to see if there's anything weird about it because we don't have anything I don't I don't think we have anything bad old that's. A lot.
I wonder I mean they wouldn't still be alive if the glaciers weren't there ran.
On may thirty first its world. No tobacco day.
This was created in nineteen eighty seven to promote awareness awareness of. The. I guess. Not just the death but you know all the. Obviously tobacco is not very good for you so unawareness of. All of those negative facts and maybe try to inspire some people to quit. Smoking. So. As a.
The latest studies. It's estimated that tobacco causes five point four million deaths every year so that's a pretty big number. It's bigger than pretty much everyone out there.
Yep five point four million deaths.
Well that's because you're eating it's not because you're. It's not because you're not smoking I mean there. Supposedly there are some cognitive benefits to to make a team. But. The costs far outweighs the benefits as. And now it's it's pretty clear at this point.
S.. I mean if they if they came out with tobacco. Today. Like it would never be approved it would never be approved there's no way.
No way and you know there's so many other. Substances and drugs that are illegal but none of them are as bad as tobacco. None of them are as harmful or not of from. Kill as many people none of them are as addictive they say that tobacco is more addictive than you know even the hardest drugs out there.
Yeah that's pretty mind blowing. So don't smoke on may thirty first.
view raw watson.txt hosted with ❤ by GitHub

定量的にオリジナルの文字起こしと比較するため、レーベンシュタイン距離からテキストの一致度を測定するFuzzyWuzzyと、意味論からセンテンスの類似度を測定するDandelion APIParalleldots AIにかけてみました。

Amazon Transcribe IBM Watson
FuzzyWuzzy 91% 90%
Dandelion API 91.57% 90.06%
Paralleldots AI 4.98/5.00 4.98/5.00

Transcribeが文字起こしの精度でWatsonに優っていることがわかります。実際に内容をみてみると、数字以上にTranscribeの実用度の高さと両製品の設計思想の差がわかります。

オリジナル Amazon Transcribe IBM Watson
I mean you can go to karaoke or you know, there’s other things to do. i mean you go cut ok that others are munitions in there. I mean you can go to cuddle care. You know there's a.
three club executives at the club Vanity in Roppongi were arrested. three club executives the club vanity and bundy were arrested Three club executives at. The the club vanity and open the gate were arrested.

カラオケ、六本木といった英文で使われる日本語固有名詞の聞き取りに両方とも失敗しています。これは今後機械学習で外来語の語彙が蓄積されれば解決できる問題かと思います。

カラオケのくだりは後半の音声が不明瞭な上にMamiの笑い声が挟まるので推測で文章を補う必要があります。オリジナルはMichaelが言いかけた others という単語をカットして文章を繋げていますが、Transcribeはそこを文字にし、(本来意図する表現ではないが)意味が通る文章にしています。Watsonは文字起こしを途中で放棄しています。

オリジナル Amazon Transcribe IBM Watson
there’s so many like really clever hackers out there. And it’s just a, it’s such a zoo. they're so many alike. Really clever hackers out there and it's just it's such a zoo. there's so many like. Really clever hackers out there. And. It's just that such as soon.
These are bryophytes, the plants. These air bridal fights the plans these are bryophytes the plants.
You know it was a very similar situation with TEPCO. you know it was a very similar situation with that guy. you know it was a very similar situation with KEPCO.

上記の3例では、製品が使っているコーパスと推測根拠がよくわかります。

it's such a zoo は、Teddy Malcolm Sousaresの楽曲にあるように、いま風の会話によく出てくる表現ですが、Watsonは聞き取りに失敗し、かつ意味の通らない文章に起こしています。一方で、コケ植物門を意味する学術用語である bryophytes を正確に聞き取っています。

韓国の原発の話題の中で、唐突に TEPCO という単語が出て来ましたが、Transcribeは認識できずに意味が通る2語への置き換えを行なっていますが、Watsonは韓国電力公社 KEPCO と誤って起しています。実際に音声を聞いても KEPCO とはとても聞こえないのですが、これは前後のキーワードから話題を推定した上で音声認識を行うIBMのCognitive Computingの仕様からおきた間違いです。

こういったキーワード推測のアプローチはクイズの問題を認識して答えを導き出すシステムとしては長けていますが、文字起こし用途やアプリケーションを作成する上では、Transcribeの意味の通る文章を作ろうとする戦術の方が向いていると思います。

Transcribeは文字起こし用途で実用的か

文字起こしの過程と種類をおおざっぱに分類すると、機械起こし、それを編集した逐語、言い間違いをそのまま表記する素起し、過剰な相槌や言い直しを修正したケバ取り、文章校正を含んだリライトがありますが、Transcribeはそのどれにも属さない、「逐語かつ意味が通らない発話を校正した」文字起こしで、極めて実用的です。

バイリンガルニュースのオリジナル文字起こしも音声情報だけでは不可能な推測による補完をしていますが、Transcribeの文字起こしは素のままでとにかく読める体裁を取っているため、そういった補完や、用途に応じた文字起こしに整形することが容易です。そのセンスは音声認識ソフトによる機械起こしから職人芸で修正と校正をタイプしていくプロセスを丸ごとカットできる技術革新といえます。出力されたJSONファイルから文字起こしアプリに入れるデータにするまで、さほど時間はかからないはずです。

先出の事例のように、ビジネスの現場に存在する膨大な音声データを迅速かつ自動でテキストマイニング可能にするほか、これもプレビューが開始されているAmazon Translateとの組み合わせで、特に報道機関の仕事のやり方が大きく変わっていくと思われます。

プレビューのお申し込みはこちらから。ぜひ一度試していただきたい今年イチオシのサービスです。

 


クラスメソッドで、ヨーロッパリージョンで勤務しませんか? クラスメソッドでは、日本、バンクーバー、ベルリンで一緒に働く仲間を募集しています! 採用情報 | クラスメソッド株式会社