Record yourself: working on my spoken French with voice recognition AI

🎙Give me a feedback loop and I'll dial down my Eastern-european accent

Here’s a dialog I’ve been in hundreds of times:

Przemek: Hello, I’m looking for a bottle of red wine, maybe something from the Loire valley?
Shopkeeper: Ah, I hear you have a little accent, where are you from?
Przemek: (blankly stares into the void)

(Transcript in English, original conversation in French.)

It took me a long while to learn French. After living in Paris for 10 years, I can finally read in French, talk in French, and even do creative things (improv!) in French... But, I still seemingly can’t lose my Eastern-european accent.

Dude

To learn anything, we need feedback loops: something that tells us what we do right and what we do wrong.

Pronunciation is hard to perfect, because once we manage to make ourselves understood, the feedback loop disappears.

No-one ever interjects to say “Hey Przemek, the way you pronounce the word XYZ sounds a bit funny.” (No-one, apart from my friend Alexander who once told me that I’m pronouncing the English word “dude” is if it was spelled “dyude”, and it’s not supposed to be like that; I’m eternally grateful 🫶.)

The shopkeeper will say “Ah you have a little accent”, but they will never point out the specific misplaced emphasis or the off sound you need to work on.

To improve my pronunciation, I need a feedback loop.

Record yourself

Back in high school I took a workshop in radio journalism. The teacher told us two things:

1. to improve your speech, you need to record yourself and listen to the recording…

2. … but it’s cringy, so nobody does it

But hey, you know what else is cringy? Hearing comments on your accent from strangers 🙃, so I guess it’s time to bite the bullet.

The method

I’m following a simple method: I take one short snippet of French per day and I work on it for 5-10 minutes.

Today I picked this 🎞️ jewel of Jacques Brel talking about his idea of talent and work.

🇬🇧 Wanting to achieve a dream is the talent. Everything else is sweat, it's perspiration, it's discipline, I'm sure of that! Art, I don't know what it is. Artists, I don't know any. The only thing that exists are people who work on something.

I open the voice recorder on my phone and I repeat after Brel. Then, I listen to the recording. Then, I listen to Brel.

Rinse, repeat.

Hearing my own voice from the outside, I notice all the little intonation differences.

And since we live in the age of AI, I can enable voice transcription in the app. Now the computer transcribes what it thinks I’m saying:

That’s another feedback loop! In my first attempt, in at least three places the model didn’t catch the word correctly.

It doesn’t take many repetitions to get better. In 5 to 10 runs, I get to something that the model understands correctly, and also sounds much closer to the original in my ears.

Conclusion

The goal is not to “lose” the accent, just to dial it down. Get the pronunciation to be a bit more fluid and natural.

The method for this has been known by generations of radio journalists and voice artists: you shall record yourself and listen to the recording.

It’s cringy, but it works 💫.

Postcard from Paris

Testing the lights. Last weekend we did an improv session in an unlikely place. The teacher reassured us it was only the 3rd weirdest place in which he was ever invited to work.

Stay warm 🍵,
– Przemek