Conversational Speech in Manipuri in the Medical Domain

Dataset creation

The dataset is strictly comprised of natural, spontaneous speech. Speakers were explicitly instructed not to follow a script. They were told to pick a topic they felt comfortable with and think: "How do I talk about this in real life?" To help them sustain natural dialogue for the required 9 to 10 minutes per session, speakers were given a tip to map out 4 to 5 simple bullet points or practice chatting with their partner once before hitting record. They were told to keep it personal, share their own stories, what they have seen around them, and what they think or feel.

The recordings feature natural pairs of individuals who already have an established real-life relationship. Users were told to record the conversations with someone they know, like a friend, a cousin, or a neighbor. To ensure acoustic and conversational diversity, a user could only record with the same partner for a maximum of 3 sessions, and they had to select an entirely different topic each time. For every recording, explicit speaker metadata was logged for the conversational partner, including Age, Gender, Phone Number, and the PIN code of where they are originally from.

Setup and Recording Instructions Given to Speakers: Speakers were instructed to talk in a quiet place so their voices remained clear. While minor background noise was acceptable, it could not interfere with clarity, and both individuals had to be equally audible using the phone’s standard built-in microphone. Speakers were given an IMPORTANT safety instruction: they were strictly forbidden from mentioning their real names, their partner's names, where they work, or where they live. All identifying personal details were intentionally kept out of the conversations. Every submission had to strictly adhere to a window of 9 to 10 minutes. Participants were required to complete a minimum of 4 and a maximum of 8 conversation recordings per assigned domain.

Directory structure

The root directory contains pairs of WAV and JSON files, aligned with a prefix:

• T-18044.MT-ee55b0648bbd4e3aa1863de8d4bdcfc0.MTA-fa9fed6ffc92421a80c19488dc9b125d.A-562950054523254.W-130571.K-user_recording.R-87ec30a3fc9f4075.merged.json

• T-18044.MT-ee55b0648bbd4e3aa1863de8d4bdcfc0.MTA-fa9fed6ffc92421a80c19488dc9b125d.A-562950054523254.W-130571.K-user_recording.R-87ec30a3fc9f4075.wav

File structure

The JSON file has contains three parts: “audio”, “metadata” and “transcriptions”:

• audio: The filename and file format information

• meta The metadata of the speaker and the topic

• transcriptions: The time aligned transcriptions

Example

{
  "audio": {
    "filename": "T-18046.MT-add4ab8841fa4b7baec244be22f1c94a.MTA-ee29aefef317429388d74a55c921e0c8.A-562950054527938.W-142282.K-user_recording.R-b60e736d2b24d959.wav",
    "duration_sec": 331.759501,
    "sample_rate_hz": 44100,
    "bit_depth": 16,
    "channels": 1
  },
  "metadata": {
    "speaker2_gender": "male",
    "user_topic": "[\"ꯅꯨꯄꯤꯁꯤꯡꯒꯤ ꯑꯃꯁꯨꯡ ꯃꯃꯥ ꯑꯣꯏꯕꯤ ꯁꯤꯡꯒꯤ ꯍꯛꯆꯥꯡꯒꯤ ꯐꯤꯕꯝII\"]",
    "speaker2_pincode": 795001,
    "workerid_gender": "male",
    "val_id": "00342da2476949c798c841e3dccefe36"
  },
  "transcriptions": [
    {
      "utt_id": "T-18046.MT-add4ab8841fa4b7baec244be22f1c94a.MTA-ee29aefef317429388d74a55c921e0c8.A-562950054527938.W-142282.K-user_recording.R-b60e736d2b24d959_0001",
      "speaker": "SPEAKER1",
      "start_time": 0.31,
      "end_time": 2.53,
      "text": " ꯏꯇꯥꯎ ꯏꯇꯥꯎ ꯀꯔꯝꯃꯥꯢ ꯇꯧꯔꯤꯒꯦ ꯍꯨꯖꯤꯛꯀꯥꯟ ꯁꯤꯕꯣ? "
    },
    {
      "utt_id": "T-18046.MT-add4ab8841fa4b7baec244be22f1c94a.MTA-ee29aefef317429388d74a55c921e0c8.A-562950054527938.W-142282.K-user_recording.R-b60e736d2b24d959_0002",
      "speaker": "SPEAKER2",
      "start_time": 2.58,
      "end_time": 10.38,
      "text": " ꯑꯁ ꯏꯇꯥꯎ ꯅꯨꯡꯉꯥꯏꯔꯤ ꯍꯥꯏꯇꯥꯔꯦꯅꯦ ꯀꯃꯥꯢ ꯀꯩꯕꯨ ꯍꯥꯏꯔꯨꯅꯤ ꯅꯨꯉꯥꯏ ꯍꯥꯏꯕ ꯅꯠꯇ꯭ꯔꯗꯤ ꯑꯗꯣ ꯏꯇꯥꯎꯁꯨ ꯎꯗꯕ ꯀꯨꯏꯔꯦ ꯏꯇꯥꯎꯕꯨ ꯏꯇꯥꯎꯁꯨ ꯅꯨꯡꯉꯥꯏꯔꯤꯕꯣ ꯀꯃꯥꯏ ꯇꯧꯔꯤꯒꯦ꯫ "
    },

    ….

Translated Examples:

1. ꯇꯁꯦꯡꯅ ꯍꯥꯏꯔꯥ ꯀꯥꯟꯗꯗꯤ ꯅꯨꯄꯤ ꯑꯣꯏꯕꯁꯤꯡꯗ ꯈꯛ ꯅꯠꯇꯅ health ꯒꯤ ꯗꯁꯨ ꯌꯥꯝꯅ ꯁꯣꯛ ꯍꯜꯂꯦ ꯂ ꯑꯩꯈꯣꯏ ꯏꯁꯥ ꯃꯛꯇ ꯇꯁꯨ ꯁꯣꯛ ꯑꯦ ꯕꯀꯣ꯫

• The pain has increased, and not only for women. My body is also suffering from this pain.

1. ꯑꯈꫭꯏꯅꯁꯨ ꯈꯪꯅꯕ ꯑꯩꯒꯤ ꯏꯃꯥꯟꯅꯕꯁꯤꯡ ꯀꫭꯂꯤꯛꯁꯤꯡ ꯃꯌꯥꯝ ꯇꯥꯟꯅꯔꯒ ꯍꯥꯏꯇꯥꯔꯦ gynaecologist ꯀꯤ ꯗꫭꯛꯇꯔ ꯑꯐꯕ ꯑꯃꯗ ꯐꯖꯅ ꯌꯦꯡꯊꫭꯛꯑꯒ ꯅꯥꯔꯤꯕ ꯃꯔꯝꯁꯤꯅ ꯀꯔꯤꯅꫭ ꯀꯃꯥꯏ ꯇꯧꯔꯒ ꯅꯥꯕꯅꫭ ꯀꯩꯗꯧꯕꯩꯅꫭ ꯈꫭꯠꯄꯩꯅꯦꯗꫭ ꯐꯖꯅ ꯊꯤꯗꫭꯛꯑꯒ ꯍꯥꯏꯇꯥꯔꯦꯅꯦꯀꫭ ꯑꯗꯨꯃꯥꯏ ꯇꯧꯔꯒ ꯌꯦꯡꯊꫭꯛꯄꯁꯤꯅꯗꯤ ꯌꯥꯝ ꯐꯩ ꯍꯥꯏꯇꯥꯔꯦ ꯑꯩꯒꯤ ꯃꫭꯠꯇꯗꯤ꯫

   • After consulting my friends and colleagues and also a good gynaecologist, why there is illness and how it started suffering there needs to be fully checked, this will be very good for us in my opinion.

2. ꯏꯇꯥꯎꯒꯤ ꯈꯪꯅꯆꯥꯅꯕ ꯃꯈꯩꯗꯨ ꯍꯡꯗꯅ ꯈꯤꯠ ꯇꯅ ꯀꯔꯤ ꯇꯧꯕꯥꯌꯥꯏ ꯀꯔꯝꯕ ꯂꯩꯅꯥ ꯑꯣꯏꯔꯛꯄꯥ ꯌꯥꯏꯗꯨ ꯑꯩꯁꯨ ꯈꯜꯂꯤꯅꯤ ꯊꯥꯗ ꯈꯣꯡ-ꯀꯥꯞ [inaudible] ꯍꯥꯏꯗꯤ ꯍꯦꯟꯖꯤꯟꯅ ꯂꯥꯛꯄ ꯍꯥꯏꯁꯤꯕꯨ ꯀꯔꯤꯅꯣ ꯑꯃ ꯗꯤ ꯁꯣꯏ ꯔꯝꯃꯅꯤ ꯑꯣꯏꯔꯝꯕꯥꯌꯥꯏ ꯍꯥꯏꯅꯑꯩꯁꯤ ꯅꯍꯥꯟ ꯗꯩꯩ ꯈꯜꯂꯤ ꯕꯅꯤ ꯑꯗꯨꯗꯨꯒꯤꯅꯦ ꯑꯩꯅ ꯂꯛꯄꯧ ꯍꯟꯗꯛ ꯊꯥ ꯁꯤꯗꯁꯨ [sneezing] ꯏꯇꯥꯎꯗ ꯀꯨꯏꯁꯤꯟ ꯗꯔꯤ ꯀꯥꯟ ꯗꯥꯁꯨꯀꯣ ꯑꯩ ꯏꯇꯥꯎꯗ ꯐꯠꯇ ꯕꯥ ꯍꯥꯏ ꯄꯨ ꯗꯨꯁꯨ ꯑꯣꯏ ꯍꯟ ꯅꯤꯡ ꯗꯕꯒꯤ ꯍꯌꯦꯡ ꯃꯛ ꯍꯦꯛꯇ ꯂꯥꯛ ꯑꯒꯦ꯫

   • After consulting your relatives and friends, what would be the better option, and what disease might it be, I'm also thinking about that. The monthly period [inaudible], which is supposed to be occured once in a month , it occured twice [sneezing] which is not normal, and there might be some issues with her, so I'm thinking about this situation. So before it gets worse, I'll come tomorrow soon.

Speakers

Multi-speaker dialogue setup containing contributions from distinct native speakers, domain professionals (healthcare assistants), and local consumers.