This project sought to create an anki deck of minimal pairs for vowels. To download the anki deck, please click this link: French Minimal Pairs for Vowels.
I am going to use forvo's api ($12 for six months, a little overpriced to use once) in order to generate an anki deck of minimal pairs to be able to identify French words.
The initial idea came from fluent forever's minimal pairs anki decks (which I have not purchased). After I read this incredibly well-written review of the product I decided it wasn't the right choice for me.
Since I didn't want to spend the $12 on his deck, but I did want to explore the concept, I found this pretty good anki deck for French pronunciation. It has 349 notes, and is well put together, but I feel it's not enough, specifically given this quote from that review I linked:
Perhaps a better approach would have been to have only recorded people speaking their own language but to have a large enough pool of voices that it would be hard to recognise a particular voice. That would also help with the dialect problem because you could either have words pronounced in each major dialect or at least consistently hear the same dialect. I understand that having the extra speakers would have cost more but then much of the money raised was used for producing word lists, cartoons and free copies of the book, none of which contribute to the core purpose - improving pronunciation.
This is a brilliant idea, I thought, and it's a shame that Gabriel Wyner wasted everyone's money on pretty pictures and word lists. But then I had a thought.
My Google Fu was feeling quite tingly, so I googled French minimal pairs and found this excellent linguistics Stack Exchange answer which lead me to this blog post where a brilliant man describes developing this incredible list of every minimal pair in the French language. (He also did it for American English, Dutch, and Spanish if anyone is interested.)
The problem with this, of course, is that his list does not have audio like I want it, and it's not even a machine readable format (it's html). Tant pis. Instead of emailing him for it (which would be obnoxious), I figured I'd just break out BeautifulSoup4 and parse it myself into 4 fields:
- The 1st word
- The 1st word's IPA
- The 2nd word
- The 2nd word's IPA
This went off without a hitch.
The next step is to get all of the audio files. For $12 (at six months) I have a 500 requests per day rate limit. What's worse is that each audio file's time to be downloaded is limited to two hours (they're one time links, I guess). And, finally, each audio file download also counts toward my requests. There are 3792 pairs, which is 7584 words, and for any word there may be 4-5 pronunciations. If we average at 3 that's 22,752 downloads, which does not include the requests to get the indexes of files per word (which will require 7,584 requests), which ends up being 30,336 requests. That will take me 2 months to download.
I'm going to restart, and create a list of only the minimal pairs I care about (since I know the difference between n and m, for instance).
It's starting to feel like I would have been better off just scraping like a jerk than buying the api.
Alright, so I recognized first that I don't really need to worry about consonants. I saved like 2 pairs from that. The second thing I realized is that there are an absurd amount of duplicates in this list, not just of pairs, but of words.
First thing I did was go through the actual web page and my csv list and deleted most of the consonant pairings. That brought me down to 1487 from 3792. Since French vowels are the most difficult to hear, I'll just stick with those. And since I don't feel like figuring out which vowels are problematic (obviously I can differentiate opposition from imposition even in French) I'll just keep them all. Because...
The second thing I did was deduplicate the list of pairs. That brought me from 1487 down to 745. This is starting to sound more doable.
In those 745 pairs are 745 x 2 =
1,490 words. But there an absurd amount of
duplicates. Bringing that down to unique we have a mere 497 unique words.
497 words times an average of 3 files per word comes to 1,491 files to download. That's almost 2,000 requests which I can do in (probably) a little over 4 days.
Now I just have to write a program that will do the following:
- Take in the list of unique words
- query forvo for the word
- Make sure the word given matches the word I searched for (forvo will serve
à
even when I just searcha
, so I have to make an explicit match in Python). - Download each of the French pronunciations of the given word.
- Add the filenames to a json file to identify each word (and keep track of what has already been downloaded).
- Repeat steps 3, 4, and 5 until I hit the 500 request rate limit.
- Fail out by writing the list of dictionaries to the json file mentioned in item 5 (I decided to write it once at failtime instead of continuously write for every new word because... Well, I'm not stupid.
Then, after that, I will use combinatronics to match every wordpair with every file so that I have the maximum number of anki cards with the maximum number of sounds for ultimate learning goodness.
Now to begin this hard part.
This is turning out easier than I expected. I wrote the program (download.py) and already maxed out my first day quota. I finished 132 of the 497 words, which resulted in 356 files. A little under 3 per word.
The quota resets 22:00 UTC, which is literally in an hour, so I'll come back here this evening and run it again today.
By Tuesday, I'd say, I'll have them all.
So as predicted I finished today. Added 2 ipy files because my first attempt to create the deck was a little short sighted and I didn't think about how to cite the file name to play the audio.
Anyway, shared the anki deck and now it's linked at the top.
I wanted to give a bit of an update about the quality of the deck some months later. I have actually stopped using the deck. There are quality issues.
First, if you are going to try to create your own deck (for whatever language) pay attention to the quality of the sound files. There is a rating (votes) and a ranking to each voice selection that is provided by forvo. Use it. I didn't and the result was sometimes words that are clearly mispronounced for whatever reason.
Second, trim the sound comparisons. You do not need even half of the vowels in French. Ultimately you can probably trim the thing down to at most 5 difficult sound comparisons (/u/ vs /eu/, for instance, instead of including /u/ vs /a/ which is obvious).
For all these reasons I have more or less given up on the deck (plus it was an insane time suck because I gave myself way too many comparisons a day in the beginning). I may eventually return to this project to recreate it, because I still do have trouble distinguishing between /u/, /eu/ and a few other phonemes. I would create a deck, in that event, of just those five or ten comparisons that I have trouble with.
Regardless, I hope this repo will help someone eventually.