A speech recognizer that listens for sounds, not words

What it does Allosaurus takes mono WAV audio and emits a sequence of IPA phones. It ships with a pretrained universal model covering roughly 2000 languages, plus an English-specific variant if you want tighter accuracy for that one language.

The interesting bit Instead of transcribing words, it transcribes the raw sounds—like a very patient phonetics professor in software form. The underlying model was trained on a multilingual allophone system, so it attempts to map acoustics to language-independent phones before narrowing to a specific inventory if you ask it to.

Key highlights

• Supports around 2000 language inventories, though most rely on a single universal acoustic model
• Outputs standard IPA symbols, with an optional timestamp approximation for each phone
• Lets you customize the phone inventory by editing a plain text file of IPA symbols
• Can emit top-k probable phones per frame if you want ambiguity explicitly spelled out
• Runs inference on CPU by default; GPU optional

Caveats

• Timestamp accuracy is explicitly noted as approximate because of the CTC architecture
• Only one language-dependent model (English) is currently available; everything else falls back to the universal model
• Expects mono-channel WAV input; anything else needs preprocessing

Verdict Grab it if you do phonetic research, low-resource language documentation, or speech analysis where word-level recognition isn’t the goal. Skip it if you need transcripts you can actually read at a dinner party.