How IPAtics transcribes

Dictionary first, AI second. IPAtics looks every word up in a pronunciation dictionary for its language. Only when the dictionary has no answer does an AI model write the transcription. Audio comes from text-to-speech voices, and reports of wrong transcriptions feed back into what everyone sees.

For every language but one, IPAtics carries a pronunciation dictionary of its most frequent words, tens of thousands per language, chosen from the wordfreq frequency lists. Arabic's is larger and built differently, as described below. The website's IPA translator and the desktop app read the same dictionaries.

Most entries were generated by open-source pronunciation software, then cleaned up with rules written for each language:

• English (US and UK), German, French, Spanish, Italian, Brazilian Portuguese, Russian and Turkish: eSpeak NG.
• Mandarin: pinyin readings from pypinyin, converted to IPA with the tones written as Chao tone letters.
• Japanese: kana readings from pykakasi, converted to IPA.
• Korean: the g2pK pronunciation rules, converted to IPA.
• Arabic: vowelled entries from Wiktionary, as extracted by Kaikki, including their inflection tables, with CAMeL Tools adding the vowels to frequent words Wiktionary does not cover. Lookups ignore the short-vowel marks in your text, so vowelled and unvowelled spellings find the same entry.
• Hindi: no dictionary yet, so Hindi relies on the fallbacks described in section 3.

Because most entries are generated rather than copied from a published dictionary, some are wrong. We correct them as we find them, and the report button described below is the quickest way to point one out.

Transcriptions are broad and phonemic, written between slashes: one symbol per sound that distinguishes words, without the finer detail of a narrow transcription. The difference is explained in phonemic vs phonetic transcription. The exception is American English, which keeps two details of General American speech: the flapped t of water, /ˈwɔːɾɚ/, and the glottal stop of button, /ˈbʌʔn̩/.

In languages whose dictionaries mark stress, the mark ˈ goes at the start of the stressed syllable, as in /bəˈnænə/, not just before its vowel. French transcriptions carry no stress marks, because French stress falls on the last syllable of a phrase rather than belonging to a word.

Each language is transcribed in one reference variety, so a transcription shows its standard pronunciation, not every speaker's. Where that variety has a name of its own, it is:

• English (US): General American
• English (UK): Received Pronunciation
• Arabic (MSA): Modern Standard
• Portuguese (BR): Brazilian
• Spanish: Castilian

A dictionary keeps one pronunciation per spelling. Where one spelling has two pronunciations, like English read in the present and the past, it shows only one, and a sentence does not change which one.

If you have chosen a language and a word is not in its dictionary, IPAtics can fall back on letter-to-sound rules for that language. Arabic never uses them, because Arabic spelling usually leaves out the short vowels and rules alone would guess them.

If the dictionary and the rules cannot cover every word of your text, the text goes to an AI language model from OpenAI, instructed with the sound inventory of the target language and asked for a broad transcription. Its answer is cached, so asking again returns the same transcription instead of a new guess.

Going the other way, from IPA back to words, IPAtics reads the same dictionaries backwards and offers every spelling that matches. When a transcription only matches loosely, or some of its words are not in the dictionaries, the AI model is asked instead and given the closest dictionary words as hints.

Spoken audio comes from Google Cloud Text-to-Speech voices. If Google's quota is used up, the app switches to OpenAI's text-to-speech so you still hear the word. Reading a raw IPA string aloud only works with Google's voices, so that has no fallback.

The example words on the phoneme pages were recorded once with the same Google voices, in a US and a UK version for English words.

Dictation in the desktop app's Editor sends your recording to an OpenAI speech-to-text model, and the words it writes down are then transcribed like typed text.

The Speech Analyzer works differently. A phoneme-recognition model (wav2vec 2.0 XLSR-53, fine-tuned for phonemes) runs on your computer, writes down the sounds you produced and compares them with the target's transcription. The difference between the two approaches is explained in audio to IPA.

The example tables in our cheat sheets take their IPA from a second set of dictionaries, built by preferring transcriptions published in Wiktionary and similar sources over generated ones. A test fails if any table shows a transcription that differs from that dictionary.

In the desktop app, use Report wrong transcription on the overlay; saved words, history and the Editor have the same option. Say what is wrong (the sounds, the stress, the language, or all of it), and add the correct IPA or a comment if you know it.

Reports are processed in batches. A correction replaces the stored transcription for everyone only when more than one person has reported the same word, or a Premium user has supplied a valid correction, and an automated check agrees with high confidence. A report without that backing waits for others, and a case the check is unsure about is held for a person to review.

Found a mistake on this website? Email [email protected] with the word, its language and the transcription you expected.