BLP Datathon 2027 — Call for Participants

Task 3 of the BLP 2027 shared tasks.

Note: This page is an overview. We publish the full instructions and rules when the platform and registration open (section 2).

1. Read this first

The BLP Datathon 2027 asks volunteer teams to record regional Bangla speech, so that we can find where open-source speech-to-text models fail. You do not need research experience. If a word is new for you, see section 12, “Key words”, near the end.

1.1 What is the goal?

A speech-to-text (STT) model does automatic speech recognition (ASR). It takes an audio file with speech as input, and it gives the spoken words as written text. This task is very important, because it makes these tools possible: voice assistants, voice search, dictation, and captions for deaf and hard-of-hearing persons. It also lets persons who cannot read or type use phones and online services with their voice.

Most Bangla ASR models learn from standard Bangla. As a result, they probably make many more errors on regional speech, for example speech from Sylhet, Chattogram, Noakhali, Rangpur or Barishal. The datathon will measure this, with real speech from real people. When you contribute to this effort, you help speech technology work for the people of your own region. You also help to make a public dataset. Researchers everywhere can use it to build better models, and it keeps a record of regional Bangla for the future.

1.2 How does this help your career?

The datathon gives you real research work that you can show to other people. We do not promise a job, a scholarship or a recommendation letter. But each item below is a fact that you can put on a CV and explain in an interview.

Skills that last. You practice careful work with data, teamwork, research ethics and clear writing. Admission committees and employers look for these skills in all fields.

Make it count. Keep notes of what you did: the number of speakers, the dialects, and the problems that you solved. Use these facts in a statement of purpose or an interview. Your certificate shows your Member ID and your points, so a reader can check them on the final leaderboard.

1.3 What else do you get?

• Points on a public leaderboard, for your team and for you as a person.
• Your name on the public credit page, if you agree.

1.4 What does it cost?

Registration is free. You need a phone that records audio and internet access to upload files. A computer is better for transcription, but a large phone screen can also work.

2. Timeline

This table is the only place in this call with dates. Other sections use the milestone names. If a date changes, we change it here only. All deadlines are at 23:59 Bangladesh time (UTC+6).

3. Who can take part

Any person who is 18 or older can take part, from any country. You must read and write Bangla well, because you will type transcripts in Bangla script. Students from Bangladesh and West Bengal are our main audience, and no research experience is necessary.

A team has 1 to 5 members. Each person can be a member of only one team. Each team selects one team lead.

4. How it works

We publish the step-by-step instructions when the platform and registration open (section 2).

4.1 What does your team do?

Your team finds volunteer speakers and records each speaker for approximately 10 minutes. Then your team types the exact words of each recording. You do both offline, at your own time. Then you upload your sessions, and we review the data and give points for it.

4.2 What happens to the data?

We give all accepted data to a set of ASR models and measure their errors. Teams get points for data that is accepted, diverse and difficult for the models. After the datathon, we publish the dataset for research.

4.3 How much time does it take?

One session takes approximately 30 minutes with consent and setup. Its transcript takes 1 to 2 hours more. A team with 10 sessions needs approximately 15 to 25 hours of work in total.

5. How to record a session

A good session sounds like a normal talk in a normal place. We want real speech from real places, so you do not need a studio. A phone is sufficient. The exact order of a session, with the consent part, is part of the full instructions (section 2).

5.1 Equipment and settings

• Use the voice recorder app that came with your phone.
• Set the highest quality. WAV is best. M4A and MP3 are also acceptable. Most phones record at 44.1 or 48 kHz, which is correct. Do not use AMR or 3GP formats. They record at 8 kHz, and we reject them. Most Android phones (for example Xiaomi, Realme, Oppo, Vivo, Samsung, Infinix or Tecno) have a format setting in the recorder app. Open the settings of the app and select WAV, or M4A or MP3 at the highest quality. If the app has noise reduction or a “meeting” mode, turn it off.
• Make sure that the phone has sufficient free space. A 10-minute WAV file is approximately 60 MB. An M4A file is approximately 10 MB.
• Charge the phone. Set it to airplane mode or “do not disturb”, so that calls and alert sounds do not stop the session.

5.2 Place

• Record where the speaker lives or works: a home, a shop, a field or a courtyard.
• Normal sounds are acceptable, for example birds, traffic or a fan.
• If you can, turn off a TV, a radio or music near the speaker.
• Put the phone on a table or hold it still, 20 to 30 cm from the mouth of the speaker. That is approximately one hand span.
• Try to make sure when speaker speaks, no one else is speaking nearby, so that we get only one voice.

5.3 Topics for free talk

Childhood memories, local food and how to cook it, a festival in the village, farm work or daily work, a journey, the local market, seasons and weather, a folk tale. Local issues are also good topics: prices, roads, power, water, jobs, schools, hospitals and local services.

5.4 Topics to avoid

Support for or against a political party, names of living politicians or local leaders, calls to protest or to vote, and accusations against named persons. Also avoid religion, health, crime, money problems and the private life of other people. If the speaker starts one of these topics, gently go to a different topic.

5.5 The interviewer (you)

• Ask short, open questions, for example “Tell me about the weekly haat in your village.”
• Speak little. Do not interrupt the speaker.
• If possible, ask questions in the dialect of the speaker.
• Only one speaker per session. Other persons must stay silent. If another clear voice speaks, we cannot use that part.

5.6 Do not do these things

• Do not edit, cut, clean, speed up or slow down the audio.
• Do not send audio through WhatsApp, Messenger, Telegram or IMO. These apps make the audio quality worse.
• Do not record TV, radio, YouTube or any recorded voice.
• Do not use AI voices or any tool that changes a voice.
• Do not record a person who is younger than 18.

6. How to transcribe

The main rule is simple: write what you hear, not what is “correct”. A transcript that changes dialect words into standard Bangla is wrong for this datathon. We publish a full transcription guide with dialect spelling examples with the full instructions (section 2).

6.1 Keyboard

Use a Unicode Bangla keyboard, for example Avro Keyboard on a computer or Gboard (Bangla) on a phone. Do not use old ANSI fonts such as SutonnyMJ. Our tools cannot read ANSI text.

6.2 Do not use AI

Do not use any ASR, AI chat tool or translation tool to make a first draft. Type each transcript yourself. A transcript copied from a model gives a wrong measurement, and we can find it (section 8, “Rules and penalties”).

6.3 How long does it take?

Expect 6 to 10 minutes of work for 1 minute of audio. A 10-minute session takes 1 to 2 hours to transcribe. Plan your team time with this number.

7. Credit and papers

Your team can write a paper about its own work in the datathon, for example how you collected data, what you learned about your dialects, or the errors of models on your data. This is often the best first paper for a student. We share the template, the instructions and the review process later. Section 2 gives the team paper deadline.

Contributors get a certificate. A public credit page lists the contributors who agree to show their name.

8. Rules and penalties

The rules protect the speakers, keep the competition fair and keep the data useful. When you register, you agree to all of them.

8.1 Rules for people

1. All participants and all speakers must be 18 or older.
2. Speakers are volunteers. Do not pay them with money or gifts. Tea and snacks at the session are acceptable.
3. Do not record a person who depends on you, for example your student, your employee or your tenant. Their “yes” is not free.
4. A speaker can speak in sessions of one team only.
5. Each team member uses their own Member ID. Do not upload work under the name of a different person.
6. Teams can add members or join with another team until team changes close (section 2). After that date, teams cannot change.
7. Obey the law of the place where you record. Before you record outside Bangladesh and India, contact us, because other data protection laws can apply, for example the GDPR in Europe.

8.2 Rules for data

1. Each session starts with the voice consent.
2. Record only natural human speech. No TV, radio, YouTube, AI voices or voice changers.
3. Do not edit the audio in any way.
4. Write the metadata truthfully. Do not change the age band, the district or the dialect of a speaker to get more points.
5. Type each transcript yourself. Do not use ASR, AI tools or translation tools.
6. Keep all data private until the public release. Do not share it, post it or put it in a paper, except the data of your own team in your team paper.
7. Use the speakers, their contact details and the sessions only for this datathon. Do not use them for a different project or product.

8.3 Rules for conduct

Respect all persons in the datathon. The ACL anti-harassment policy applies to all participants and organizers. If you write a team paper, follow the ACL policy on AI writing tools.

8.4 Penalties

If a team breaks the rules, we can remove data, give a point penalty or disqualify the team. Before a serious penalty, we tell the team, and the team can explain.

9. Data, privacy and ethics

A voice is personal data, because it can identify a person. For this reason, we treat each recording with care, and we ask you to do the same.

9.1 Who is responsible

The BLP Datathon organizers are responsible for the data. They are a group of researchers, not a legal organization. Send all data questions and removal requests to the organizers.

9.2 What we never publish

We never publish the names of speakers, the consent parts of the recordings, phone numbers, addresses and the exact village of a speaker.

9.3 Responsible use

We ask each user of the data to follow two requests. Do not try to find the identity of a speaker. Do not copy the voice of a speaker with AI voice tools.

9.4 Keep the data safe

Keep the audio only on a device with a password, and delete it after the camera-ready date (section 2).

9.5 Embargo

Until the public release, nobody can publish or share datathon data. This includes organizers. A team can use only its own data in its team paper.

10. Tools, support and contact

All tools are free. As the datathon continues, we publish more information about the tools, the links, the support and the organizers on our website. For questions, contact the organizers.

11. Frequently asked questions

How do I explain the project to a speaker? You can say this in Bangla, or in the dialect of the speaker:

আমরা আঞ্চলিক বাংলা নিয়ে একটি গবেষণা করছি। কম্পিউটার যেন আমাদের এলাকার ভাষা বুঝতে পারে, সেজন্য আপনার কণ্ঠে ১০ মিনিট কথা রেকর্ড করতে চাই। এর জন্য কোনো টাকা দেওয়া হবে না, আর আপনি যেকোনো সময় থামতে পারবেন।

In English: “We do research on regional Bangla. We want to record 10 minutes of your voice, so that computers can understand the speech of our area. There is no payment, and you can stop at any time.”

My mobile data is limited. What can I do? Record in M4A. A 10-minute session is then approximately 10 MB. Upload on Wi-Fi when you can.

My phone records only in AMR or 3GP. Install a free recorder app that records WAV, M4A or MP3 at 44.1 kHz or more. Make a short test file first.

I made a mistake in a session. What do I do? Tell the organizers as soon as possible.

Can I take part from outside Bangladesh? Yes, if you read and write Bangla. Speakers can also live outside Bangladesh. We use the district where they grew up for the dialect zone. Before you record outside Bangladesh and India, contact us.

I work at a company. Can I take part? Yes, as a person. First, examine the rules of your employer about work outside the company. Your company gets no special access to the data.

Can a company sponsor the datathon? Yes. Contact us. Sponsors get no early access to the data, and they have no influence on scores or authorship.

12. Key words

This page uses some research words. Use this table when a word is new for you.