You have a recording: a voice memo, a clip from a podcast, an MP3 of yourself reading a paragraph. You want it in IPA. You search for "audio to IPA" and find tools that promise exactly that, and then the results look suspiciously tidy, as if nobody had an accent.
That is because "audio to IPA" means two very different jobs, and most tools do the easier one. This guide explains both, what each can and cannot tell you, and how to get the transcription you actually need.
Two different things called "audio to IPA"
The first job is recognising the words, then looking up how they are pronounced. Speech-to-text turns the audio into ordinary writing, and a text-to-IPA converter turns the writing into a transcription. The result is the dictionary pronunciation of the words that were recognised.
The second job is recognising the sounds themselves. A phone recognizer is a speech model trained to output phonetic symbols directly from audio, with no spelling in between. The result is an estimate of what the speaker actually said.
| Words, then dictionary IPA | Phone recognition | |
|---|---|---|
| What it listens for | words | individual speech sounds |
| What you get | the reference pronunciation of each recognised word | a string of symbols for the sounds it heard |
| Shows the speaker's accent? | no | partly |
| Shows a mispronunciation? | no, unless it changes the word | often, though not reliably |
| Kind of transcription | broad, phonemic | closer to phonetic, within the model's symbol set |
If those terms are new, phonemic vs phonetic transcription explains the difference between broad and narrow; it is the heart of this question.
Path 1: speech to text, then text to IPA
This is what most "speech to IPA" and "voice to IPA" tools do, and for many purposes it is the right answer. Word recognition is mature, the IPA comes from a pronunciation dictionary rather than a guess, and the transcription is clean and consistent.
What it cannot do is tell you anything the spelling doesn't. Two speakers with completely different accents produce the same text, so they get the same IPA. If a learner says ship with the vowel of sheep, a word recognizer will often still write "ship", because that is the likelier word in context, and the IPA will show a perfect /ʃɪp/. Errors also travel down the chain: a misheard word becomes a confidently wrong transcription, and a word spelled one way but said two ways, like read in "I read it yesterday", needs the converter to pick the right one.
Speech recognizers built for general use, such as OpenAI's Whisper, are trained to write ordinary text, not phonetic symbols. They are the first half of this path, not a way around it.
Path 2: recognising the sounds
Phone recognition skips the words. Two openly available examples:
- Allosaurus, from Carnegie Mellon University and collaborators, is described by its authors as a pretrained universal phone recognizer for more than 2,000 languages.
- wav2vec 2.0 XLSR-53, fine-tuned for phoneme recognition by researchers at Facebook AI (now Meta), outputs phonetic labels in many languages. Its training labels came from the eSpeak phonemizer, so its symbols follow eSpeak's inventory.
These models really do listen to the sounds, which is why they can catch an accent or a mispronounced vowel that word recognition would smooth over. But treat their output as a first draft:
- The symbol set is the model's, not the IPA's. A model can only write symbols it was trained on, so a detail outside that set, such as an unreleased stop or a particular vowel quality, is mapped to the nearest symbol it knows.
- Errors are common. Recognising individual sounds is much harder than recognising words, because there is no vocabulary to fall back on. Background noise, music, fast or overlapping speech make it worse.
- Stress and boundaries are usually missing. You get a stream of sounds, often without stress marks or word breaks.
When precision matters, phoneticians still transcribe by ear, slowing the audio down and checking spectrograms in a tool such as Praat. A forced aligner, such as the Montreal Forced Aligner, is a different tool again: given the audio and its transcript, it finds where each dictionary sound starts and ends. That gives you timing, but the symbols still come from the dictionary.
What IPAtics does with speech
IPAtics uses both paths, for two different jobs, and neither of them is a general "upload an MP3" converter. Both record from your microphone, one clip at a time, up to 15 seconds each.
Dictation, in the desktop app's Editor, is path 1. You say a word you heard but can't spell, an OpenAI speech-to-text model writes it down, and IPAtics transcribes that text like anything you type: from its pronunciation dictionaries first, with an AI model filling in words the dictionaries don't have. You get the dictionary pronunciation of what was recognised, not a record of how you said it.
The Speech Analyzer is path 2, with a target. You type the word or sentence you mean to say and read it aloud. A phone recognizer based on the wav2vec 2.0 model above, running on your own computer, writes down the sounds it heard; the analyzer shows them as "You said" next to the target's IPA and marks every target sound as correct, needing work or uncertain. It is built to check your pronunciation of a known text, which is what makes its comparison meaningful. It is not a transcriber for recordings of other people. The Speech Analyzer page explains how it works and what it needs.
How the dictionary transcriptions are made, and what happens when a word is missing, is on our methodology page.
How to get IPA from a recording today
- Get the words. Run the audio through any speech-to-text tool, or write it down by ear. Check the text: every error here becomes an error in the IPA.
- Get the broad transcription. Paste the text into the IPA translator. You now have the dictionary pronunciation, which is what most learners and teachers need.
- Narrow it by ear, if you need the speaker's version. Play the audio slowly and edit the transcription where the speaker differs: a dropped sound, a different vowel, a glottal stop for a t. A phone recognizer's output can be a starting point here, but listen to every change before you trust it.
- Check it by listening back. Our IPA to speech page covers the opposite direction: hearing a transcription read aloud.
Frequently Asked Questions
Can I convert an MP3 to IPA?
Yes, in two steps: convert the audio to text with a speech-to-text tool, then convert the text to IPA. That gives you the dictionary pronunciation of the words. If you need the speaker's own pronunciation, use a phone recognizer as a draft and correct it by ear.
Is there a voice to IPA app?
Most "voice to IPA" apps recognise your words and then show their dictionary IPA, so they will not reveal your accent. To see the sounds you actually produced, you need phone recognition against a target text, which is what pronunciation tools like the IPAtics Speech Analyzer do.
Does Whisper output IPA?
No. Whisper is a speech recognizer trained to write ordinary text. You can pass its output to a text-to-IPA converter, which gives you the dictionary pronunciation of the recognised words.
How accurate is automatic phonetic transcription?
Less accurate than word recognition, because each sound has to be identified without help from a vocabulary. It works best on clear, close-microphone speech in a language the model was trained on, and its symbols are limited to the set it learned. Treat it as a draft.
Trying It Yourself
If you already have the words, the fastest route to IPA is to paste them into the IPA translator in your browser. For your own pronunciation, IPAtics compares what you said with the target, sound by sound.
Sources: Allosaurus is described in Li et al., Universal Phone Recognition with a Multilingual Allophone System (ICASSP 2020); the wav2vec 2.0 phoneme model in Xu, Baevski and Auli, Simple and Effective Zero-shot Cross-lingual Phoneme Recognition (2021), and on its model card; Whisper in Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (2022); forced alignment in McAuliffe et al., Montreal Forced Aligner (Interspeech 2017); Praat at fon.hum.uva.nl/praat.
Related reading: Phonemic vs phonetic transcription · What is IPA? Why phonetic transcription matters · How to read IPA — a beginner's guide