Medical Domain Mexican Spanish Speech Corpus (MEDMEX)

Medical Domain Mexican Spanish Speech Corpus (MEDMEX)

10.5 hours (~5.25 hours each from two speakers, one male one female) of read-speech in the medical domain (drug/medical equipment descriptions, including name, amount, and mode of application) by licensed nurses.

Motivation

Speech technologies, particularly apart from those for English, often struggle with domain-specific content. One particularly valuable domain that speech technology systems may struggle with is medicine.

The present dataset was created with two main motivations:

1. Offer a specialized medical-domain speech corpus of Mexican Spanish for the evaluation of ASR systems, and

2. To provide specialized speech data for training a domain-specific TTS system.

Speakers

The two speakers represented in the data are both trained and practicing nurses from the state of Puebla, Mexico. They are monolingual Spanish speakers.

Dataset creation process

Text: The source text is from Mexico's National Platform for Open Data, specifically: https://www.datos.gob.mx/dataset/medicamentos_productos_farmaceuticos_adquiridos_hgmgg, "Medicines requested by the General Hospital Dr. Manuel Gea Gonzalez, desegregated by description, unit, requested amount, and destination". Our text corpus includes the descriptions and the requested amount for each drug. We publish audio for 2,623 rows, containing of 60k tokens.

Audio: Two nursing professionals were recruited to read descriptions of medication, including drug and medical equipment names, form of administration or material, and quantities/dosages. Abbreviations are read out in full (e.g. "miligramos" instead of "mg"). Speakers were instructed to practice each sentence until they could read it comfortably and naturally, as if they were reading the descriptions to a doctor or patient.

In total, each speaker contributed approximately 5 hours of recorded speech. Recordings were made in a web application and recorded in webmedia format (.webm), and later converted to .mp3, using usb microphones.

Text preprocessing

Most descriptions come in the form drug name, quantity, and mode of application. We attempted to separate each of these fields when present by a pipe character | to assist readers in pacing. The pipe characters are omitted in the final dataset.

• Cánulas Orofaríngeas De plástico transparente o translucido Tipo: guedel/berman. Tamaño: 5 Longitud: 100 mm. Pieza

• Espaciadores de volumen De plástico rígido resistente que se adapte a los diferentes medicamentos broncodilatadores en aerosol Puede tener o no ensamblada una mascarilla o una boquilla. Vida útil: tres meses. Adulto. Pieza.

• Irbesartán-hidroclorotiazida Tableta Cada Tableta contiene: Irbesartán 300 mg Hidroclorotiazida 12.5 mg Envase con 28 Tabletas.

Structure and format

The dataset directory has the following structure:

metadata.tsv
mp3/
- {key1}.mp3
- {key2}.mp3
- ...

the metadata.tsv file has the following columns:

• key: The ID of the audio file. To get the audio path from the key, prepend mp3/ and append .mp3 to the key.

• sentence: The sentence text

• speaker_sex: "Male" or "Female"

• speaker_id: A unique identifier for each speaker (1 or 2)

• duration: Length of audio clip in seconds