Perch publishes three open releases for checking drug names in dictation

Today we are publishing three open releases: a tool that checks a drug list against the audio of a dictation, a table of how veterinary drug names are said and misheard, and six blind dictation corpora.

The phoneme lane

The phoneme lane scores the audio, not the transcript. It runs a phoneme model over a dictation and scores every name on a formulary, a clinic's drug list, against the sounds it heard. It returns the best fits and the seconds of the clip where they sit. On clip r001 in blind6, read by the voice am_echo, the transcript spells famotidine "famitidine," maropitant "meripitin" and Entyce "Entice." The lane's four best fits are famotidine, morphine, Entyce and maropitant, and all four were said. In a Hindi sample sentence naming Heartgard and Bravecto, the transcript came out entirely in Devanagari, so a text matcher had no Latin letters to read. The lane ranked Heartgard first. It proposes names. It does not chart them.

Vet drug pronunciations

Vet drug pronunciations lists 265 drug and brand names used in veterinary practice, each with its pronunciation in IPA, the International Phonetic Alphabet, for American, British and South Asian English speakers. Sixty six of the names are on Perch's default list. For 59 of those the garble atlas lists the wrong words speech recognizers wrote when Kokoro voices said the name. Each is counted, and flagged if it is an ordinary English word. Atenolol, for example, came out as "8 nilal."

Blind dictation corpora

Perch blind dictation corpora are six sets of 70 rows, each a dictation or a trap, each written by an author who never saw Perch's matching code. In every set 45 are dictations and 25 are traps, sentences built to bait a system into offering a drug nobody said. Three Kokoro voices read each of the 420 rows, which makes 1,260 clips and about five hours of audio. The first set, blind1, holds the 70 clips of the trap set and adds two more voices for each sentence.

Why Perch publishes its tests

A company's own results are a claim until someone else can rerun them. The trap set post set the pattern: publish the clips and the scoring rule so anyone can check the work. These releases do the same with the sentences we score Perch on, the table our matcher reads when a name comes out garbled, and the lane we built for the names the matcher still missed.

Method notes

• The phoneme model is wav2vec2 xlsr 53 espeak, with 388 phone symbols and 20 millisecond frames.
• Every transcript quoted is Whisper large v3. The atlas holds what Whisper large v3 and MedASR wrote when six Kokoro voices said each of the 66 names in three sentence frames.
• The pronunciations are espeak IPA for American and British English, plus three spoken forms per name from a language model, one each for American, British and South Asian English.

Where to find them

• Phoneme lane, code under Apache 2.0: GitHub. DOI pending.
• Vet drug pronunciations, data under CC BY 4.0: GitHub and Hugging Face. DOI pending.
• Blind dictation corpora, data under CC BY 4.0: GitHub and Hugging Face. DOI pending.

Kokoro v1.0, from hexgrad and under Apache 2.0, voiced the clips. Run your own system against these sets, and open an issue on any of the repositories with what it misses. Perch is at perch.talcoe.com.