Over the years, CCC has supported speech data collection, machine translation post-editing (MTPE), and linguistic data processing projects involving hundreds of contributors and millions of words. Through these experiences, we have learned valuable lessons about recruitment, production planning, quality assurance, communication, and the challenges of managing large teams.

In this article, we share ten practical lessons that continue to shape how we approach multilingual AI data projects today.

Managing large-scale multilingual AI data projects involving hundreds of contributors requires more than simply having enough people to complete the required volume. Whether the project involves multilingual speech data collection, machine translation post-editing (MTPE), transcription, or linguistic quality assurance, there are many factors to consider beyond the actual production work. These include recruitment, resource allocation, quality control, communication, information security, and the ability to respond to unexpected challenges throughout the project.

At Creative Connections & Commons Inc. (CCC), we have worked on multilingual AI data projects of varying scales over the years, including projects involving millions of words of text and hundreds of speech recording participants. Our experience in translation, localization, and project management has provided us with a foundation for organizing these operations, although every project continues to present its own requirements and challenges. Through these experiences, we have gradually developed and refined our approaches to recruitment, production planning, quality assurance, and contributor management.

Some of the large-scale multilingual data projects we have supported include:

  • Tagalog Corpus Translation and MTPE (2018–2019): Approximately 100 million words processed with the involvement of around 100 contributors.
  • Multilingual Parallel Corpora – Russian (2022–2023): Approximately 120 million words involving machine translation post-editing, conversational text translation, and quality assurance.
  • Conversational Speech Corpus Localization – Tamil and Bengali (2025–2026): Approximately 40 million words involving text refinement, localization, quality assurance, and evaluation.
  • Multilingual Speech Recording and Media Annotation – Italian, French, and Singaporean English (2020–2023): More than 1,000 video hours involving transcription, automatic transcript post-editing, speaker tagging, noise labeling, and quality assurance.
  • End-to-End Scripted and Non-Scripted Speech Datasets – Tagalog and Japanese (2020–2024): More than 300 hours of recordings from over 400 speakers and script adaptation, covering business presentations, mock meetings, customer service speech, and generic conversational audio, followed by segmentation, transcription, and linguistic quality assurance.

Although these projects involved different languages, technical specifications, and delivery requirements, we encountered several common challenges in managing large contributor teams while maintaining consistent quality. Some required us to rethink our initial approaches, while others reinforced the importance of procedures we had already established through previous operational experience. In this article, we share ten lessons that have helped us improve how we plan and manage large-scale multilingual data collection and processing projects.

1. Start Small: A Pilot Phase Can Save the Entire Project

One of the most important speech data collection best practices we have learned is the value of conducting a pilot phase before committing to full-scale production. Project requirements may appear straightforward during the initial discussions, but certain challenges only become apparent once the actual work begins. Starting with a smaller volume allows both the client and the service provider to assess whether the proposed workflow is realistic and whether the expected quality can be consistently achieved.

For speech data collection, a pilot can help identify issues involving recording environments, audio specifications, participant eligibility, or transcription accuracy. These considerations are particularly important when developing custom speech datasets, where the client may require specific languages, accents, recording conditions, or technical formats. In MTPE and linguistic data processing, a pilot may reveal differences in how contributors interpret the guidelines, evaluate machine-translated content, or apply editing standards.

These observations also help us assess the resources and time required for quality assurance, which may not be immediately apparent from the production volume alone. For instance, a recording task that appears relatively simple may require several additional steps to verify audio quality, speaker eligibility, and transcription accuracy. Understanding these requirements during the pilot helps us prepare a more realistic production plan for the succeeding phases.

We believe that identifying these issues early is far better than discovering them after hundreds of contributors have already started working. A pilot provides an opportunity to clarify instructions, establish acceptance criteria, test the tools, and make necessary adjustments before investing more resources into production.

Our takeaway: We prefer starting with a manageable volume, evaluating the results, and adjusting the workflow before scaling. A successful pilot can help both parties establish realistic expectations and reduce the risk of larger problems later in production. It also provides a more reliable basis for estimating the resources needed for the full project.


2. Do Not Overcommit: Be Honest About Your Production Capacity

Large-scale data projects often come with ambitious production targets and tight deadlines, particularly when clients need substantial volumes within a limited period. While there may be pressure to demonstrate that a company can handle the entire scope, we believe that being realistic and transparent about production capacity is essential. Committing to a volume that cannot be consistently delivered may eventually compromise quality, strain the production team, and affect the client’s own schedules.

One important consideration is that the number of available contributors does not necessarily translate into equivalent production capacity. For example, a multilingual speech data collection project may require 500 speakers, but not every recruited applicant will meet the eligibility criteria, complete the recording process, or pass the quality checks. Similarly, having 100 available translators or post-editors does not automatically mean that all of them can produce the same volume or maintain the same level of quality.

When assessing a project, we consider our existing resources, experience from previous projects, recruitment requirements, expected productivity, quality assurance capacity, and possible rejection rates. If the requested volume exceeds our immediately available resources, we would rather propose a realistic starting capacity, phased production schedule, or longer delivery period. This allows us to establish a workable arrangement while giving the client an opportunity to evaluate our performance before increasing the scope.

Our takeaway: We believe that realistic commitments and honest communication are important foundations of a successful outsourcing partnership. It is better to agree on a manageable scope and deliver consistently than to overpromise and struggle to meet expectations later. A smaller but reliable starting capacity can also provide a foundation for future expansion.

3. Once Committed, Maximize Your Recruitment Channels

Once the project scope and production targets have been agreed upon, contributor recruitment becomes one of the most important operational priorities. For projects involving hundreds of participants, including crowdsourced data collection initiatives, relying on a single recruitment platform is rarely sufficient, especially when the requirements involve specific languages, qualifications, or demographic profiles. We have learned to maximize the available recruitment channels while ensuring that the screening and onboarding processes can keep up with the number of applicants.

Depending on the project, we use platforms such as LinkedIn, freelance job websites, professional networks, referrals, and personal connections. Local networks can be particularly helpful for speech data collection and voice data collection projects involving regional languages, accents, or location-specific requirements. We also encourage our existing contributors to refer qualified applicants, although these referrals still undergo the necessary screening and qualification procedures.

One example is our Tagalog speech data collection project in 2022, which involved more than 360 qualified speakers across several locations in Luzon, including Pampanga, Tarlac, Cavite, and Quezon City. The project required individual, two-person, and three-person recording arrangements covering business presentations, mock meetings, and conversational discussions. Coordinating these activities involved not only finding participants but also confirming their eligibility, explaining the recording guidelines, arranging schedules, and monitoring completion.

The resulting Tagalog speech dataset comprised more than 180 hours of recordings, accompanied by transcription and quality assurance work. The original business scenarios were provided in Japanese and underwent English machine translation and post-editing (MTPE) before being translated and adapted into Filipino to make the recording materials sound as natural as possible in the local context. This project demonstrated how multilingual data collection can involve several interconnected linguistic and operational processes before the final recordings are delivered.

We have also learned the importance of maintaining a pool of qualified backup contributors, as some applicants may become unavailable or withdraw before completing their assignments. Recruitment should therefore be treated as an ongoing process rather than a one-time activity completed at the start of production. This is particularly important for longer projects, where contributor availability and production requirements may change over time.

Our takeaway: Once we commit to a project, we make full use of our recruitment channels while maintaining consistent screening and onboarding procedures. A larger applicant pool provides more flexibility, but the objective remains to recruit qualified contributors who can reliably complete the required work. Maintaining backup resources also helps us respond to changes in contributor availability without unnecessarily disrupting production.

4. Verify Your Contributors: Be Careful with Fraudulent Applications

One challenge we have increasingly encountered in remote recruitment is verifying whether applicants are who they claim to be. With the growing availability of AI-generated content and other online tools, it has become easier to prepare convincing résumés, work samples, and professional profiles, even when the information provided is inaccurate or misleading. Unfortunately, this also creates additional challenges for companies recruiting large numbers of remote contributors.

In our recent experience, one particularly concerning warning sign has been applicants who could not provide a bank account under the same name indicated in their application documents. While there may be legitimate explanations for such discrepancies, we consider them sufficient grounds for further verification before proceeding with onboarding or payment arrangements. This is especially important when contributors will be given access to client materials, confidential information, or project-specific systems.

We have learned not to rely solely on résumés or online profiles when evaluating applicants. Depending on the project requirements, our screening process may involve identity and qualification checks, language assessments, sample tasks, live interviews (either through video calls or in person), and verification of payment details. These procedures help us confirm applicants’ identities, assess their qualifications, and determine whether they are suitable for the assigned work.

For multilingual AI training data projects, insufficient screening can affect more than the administrative side of operations. Contributors who misrepresent their qualifications or identities may compromise the authenticity, reliability, and quality of the collected data. At the same time, we recognize the importance of conducting verification fairly, protecting applicants’ personal information, and allowing reasonable opportunities to clarify discrepancies.

Our takeaway: Contributor verification should be incorporated into the recruitment process rather than treated as an afterthought. Establishing clear and consistent procedures helps protect both the company and its clients while maintaining the integrity of the project. We believe that a careful screening process is particularly important when managing large numbers of remote contributors.

5. Set Up the Infrastructure Before Scaling Production

Managing hundreds of contributors requires a clear organizational structure, particularly when production involves several stages and different quality requirements. Before starting a large project, we assess the expected production volume against our experience from previous assignments to determine the resources needed for coordination, quality assurance, and administrative support. This helps us establish the appropriate number of project coordinators and identify who will be responsible for monitoring each aspect of the operation.

For larger projects, we assign a project manager to oversee the overall production, including schedules, contributor management, quality assurance, and client communication. If the project becomes too large for a single team to manage efficiently, we divide the contributors into smaller groups with designated coordinators or team leaders. This allows responsibilities to be distributed more effectively while maintaining centralized oversight of the project’s progress and requirements.

We first applied this team-based approach during our Tagalog corpus translation and MTPE project from 2018 to 2019, which involved approximately 100 million words and around 100 contributors. Dividing the contributors into smaller teams helped us organize assignments, distribute coordination responsibilities, and monitor progress while maintaining oversight of the overall production. The experience also provided us with a practical reference for estimating coordination requirements and managing larger volumes of multilingual text data.

We subsequently built on this experience during our Russian multilingual parallel corpora project from April 2022 to February 2023, which involved approximately 120 million words of machine translation post-editing, conversational text translation, and quality assurance. Although the language and project requirements were different, the management structure developed through our earlier Tagalog project provided a useful foundation for organizing the production teams. It also allowed us to apply lessons from previous operations while making adjustments based on the specific requirements and scale of the Russian project.

The infrastructure required for speech data collection can be more complex, as the workflow may involve recruitment, recording coordination, audio processing, transcription, and linguistic quality assurance. For our 2022 Tagalog speech collection project involving more than 360 speakers, a dedicated team of six in-house staff supported the linguistic QA process using Filipino dictionary references and a client-provided checking tool. This arrangement allowed different stages of the production process to be handled by personnel with clearly assigned responsibilities.

Our takeaway: We believe that the management structure should be established before scaling the workforce. Clear responsibilities, dedicated coordinators, and appropriate quality assurance resources make large-scale operations easier to monitor and manage. Previous project experience also provides a useful reference for planning the resources required for new assignments.

6. Create a Gantt Chart and Monitor Progress Daily

Large-scale multilingual projects require continuous monitoring, especially when several production activities are happening simultaneously. We have found Gantt charts particularly useful for organizing milestones, visualizing overlapping tasks, and monitoring whether the project is progressing according to schedule. Over time, we have also learned to develop our own scripts and monitoring tools to track daily, weekly, and monthly progress, making production data easier for team members to access and review.

These tools have helped us organize production data into formats that can be easily understood by project managers, coordinators, and other relevant team members. Rather than relying entirely on manual updates or individual reports, having a shared view of production progress makes it easier to assess how each contributor or team is performing against the established targets. It also helps us identify areas requiring additional attention and make adjustments based on actual output.

A project timeline should account for recruitment, contributor screening, onboarding, training, production, quality assurance, revisions, and final delivery. For speech data collection, recording and transcription activities may overlap, with one group of participants completing recordings while another team reviews audio quality or processes previously submitted files. If progress is checked only at major milestones, delays or quality issues may have already accumulated by the time they are identified.

For this reason, we prefer monitoring production progress daily, particularly during periods involving frequent deliveries or a large number of active contributors. This allows our coordinators to compare actual output against the planned targets, identify bottlenecks, and make necessary adjustments to staffing or workloads. Daily monitoring also provides more reliable information when communicating project status or potential delivery concerns to clients.

Our takeaway: A production schedule should be treated as an active management tool rather than a document prepared only at the beginning of the project. Developing accessible monitoring tools and regularly reviewing production data helps us identify problems earlier, coordinate teams more effectively, and make informed decisions before delays affect the overall timeline. We have found that providing everyone with a convenient way to monitor progress also improves coordination and accountability.

7. Take Quality Assurance Seriously, Even When It Means Removing Contributors

One of the greatest challenges in managing large-scale multilingual projects is maintaining consistent quality across many contributors. Different people naturally have different levels of experience, interpretations of instructions, and approaches to completing their assignments. Without clear standards and proper quality control, these differences can result in inconsistent outputs, unnecessary revisions, and additional workload for the QA team.

We believe that quality requirements should be established and communicated before production begins. If the client provides detailed acceptance criteria, these should be incorporated into our internal guidelines and checking processes. When certain requirements are not explicitly defined, we may establish additional internal evaluation standards based on the nature of the work and the quality expected from the final deliverables.

For example, in one of our previous MTPE (Machine Translation Post-Editing) projects, we established a 70% similarity threshold (original MT version vs. human post-edited version) as an internal indicator for identifying outputs that might require further review. The initial machine translation quality was already relatively good, so we wanted to ensure that contributors were genuinely reviewing and refining the content rather than simply accepting machine-generated text without sufficient evaluation. When an output exceeded the threshold, it was flagged for closer checking to determine whether the level of human intervention was appropriate.

Of course, similarity alone does not necessarily indicate poor-quality MTPE, as some machine-translated passages may already be accurate and require minimal changes. For this reason, the threshold was used as a screening tool rather than an automatic basis for rejection. The actual evaluation still required human judgment, taking into account the accuracy, naturalness, consistency, and overall quality of the final text.

Speech dataset quality assurance presents additional considerations beyond those involved in text processing. Depending on the project requirements, we may need to assess audio quality, transcription accuracy, speaker information, and compliance with the recording specifications. These checks are particularly important when collecting speech data for AI systems, where inconsistencies in the training material may affect its suitability for the intended application.

We have also learned not to hesitate in addressing recurring quality issues among contributors. When problems are identified, we communicate the findings, clarify expectations, and provide opportunities for correction where appropriate. However, if a contributor repeatedly fails to meet the agreed standards despite feedback, we may need to remove them from the production line to protect the overall quality of the project.

Our takeaway: Quality assurance should be integrated throughout production, not performed only before final delivery. Clear standards, consistent checking, and timely corrective action help prevent individual issues from affecting the work of the entire team. We believe that addressing quality concerns early is more effective than allowing them to accumulate until the final checking stage.

8. Communication Is Key, Both with Clients and Contributors

In a project involving dozens or hundreds of contributors, even a small misunderstanding can affect a significant volume of work. This is why we consider communication one of the most important responsibilities of project management, both in dealing with clients and coordinating the production team. Clear and timely information sharing helps prevent unnecessary revisions, delays, and inconsistent interpretations of project requirements.

Whenever we receive new instructions or clarifications from a client, we make it a priority to relay the information to the relevant team members as soon as possible. Delayed communication may result in contributors continuing to work using outdated guidelines, potentially affecting an entire production batch. The same principle applies to delivery schedules, as we believe that anticipated delays should be communicated proactively, preferably at least one day before the scheduled delivery rather than after the deadline has already passed.

Communication with contributors is equally important, particularly for projects running over several months. For long-term projects, we often prefer establishing a common group chat where announcements, deadlines, reminders, and general clarifications can be shared with everyone. Aside from making coordination more efficient, we have found that these group chats can also encourage a sense of community among contributors.

We also use these group chats to welcome and introduce new contributors as they join the project, share and celebrate important milestones, and recognize the team’s progress along the way. At the end of a project, we make it a point to express our appreciation to everyone who contributed to its completion, recognizing the collective effort involved in delivering the final output. We believe these simple gestures are particularly meaningful in remote projects, where contributors may work together for months without ever meeting in person.

One particularly positive observation is that contributors often begin helping one another, even when they have never met in person. They share clarifications, answer common questions, and offer practical suggestions based on their own experience with the tasks. We believe this kind of interaction helps improve team morale, reduces repeated questions to coordinators, and encourages better consistency in production, provided that all communication remains within the project’s confidentiality and information security requirements.

Our takeaway: Communication is not simply about providing instructions or status updates. It is also about maintaining trust, recognizing people’s contributions, encouraging collaboration, and ensuring that everyone understands the project’s objectives and expectations. We have found that building a supportive environment can make a meaningful difference in long-term projects involving remote contributors.

9. Mistakes Can Happen. What Matters Is How You Respond.

Even with careful planning, unexpected situations can occur during production. A contributor may suddenly resign, technical problems may interrupt operations, or a delivery may fail to meet the required standards. These situations can disrupt the original schedule and require immediate adjustments, particularly when several contributors or production stages depend on one another.

While we make every effort to prevent these problems, we also recognize that not every situation can be anticipated. What matters is how quickly and responsibly we address the issue, communicate its implications, and determine the appropriate corrective action. It is equally important to communicate these developments and corrective measures promptly to the client and the affected members of the contributor team, so that everyone understands the situation and any necessary adjustments to the workflow or schedule.

For example, if the sudden departure of a contributor affects production continuity, we need to examine whether there were sufficient backup resources or whether responsibilities could have been transferred more easily. We may also need to review how the work was documented, whether project knowledge was concentrated in a single person, and what arrangements could help prevent similar disruptions in the future. The objective is not simply to replace the contributor but to make the operation less vulnerable to unexpected changes.

The same applies to quality-related problems. If a delivery requires correction, we need to understand how the issue passed through the checking process and whether additional controls or clearer instructions could have prevented it. This requires honest evaluation, including recognizing when our own planning, communication, or procedures may have contributed to the situation.

Our takeaway: We believe that mistakes should not simply be corrected and forgotten. Understanding their causes, communicating them appropriately, and applying the lessons learned can help strengthen the project and improve how we manage future operations. Every significant issue is also an opportunity to examine whether our procedures can be improved.

10. Previous Operational Experience and ISO 27001 Practices Make a Difference

Our experience in large-scale multilingual projects has reinforced the importance of established operational procedures, proper documentation, and information security management. As part of CCC’s ISO 27001-aligned information security management practices, we have developed internal policies and procedures covering different aspects of our operations. Although these were primarily designed to manage information security risks, we have found that the underlying principles are also valuable in managing complex production projects.

One of the most useful principles is risk-based thinking, which encourages us to identify possible issues before they occur and assess their potential impact on the operation. In multilingual AI data projects, these risks may involve contributor access to confidential materials, unauthorized sharing of client information, inaccurate data, insufficient QA capacity, or disruptions to production continuity. By identifying these concerns during project planning, we can establish appropriate controls and prepare possible responses before problems arise.

Documentation is another important consideration, particularly when projects involve large teams or run for extended periods. Properly documented guidelines, responsibilities, workflows, and quality requirements make it easier to maintain consistency, onboard new contributors, and implement changes without unnecessarily disrupting production. Documentation also provides useful references when investigating quality issues, reviewing decisions, or improving the workflow for future assignments.

Beyond information security, our previous operational experience has helped us become more realistic in assessing recruitment capacity, contributor productivity, human resource requirements, and project timelines. Managing large-scale assignments over the years has given us practical references for determining how many coordinators may be needed, how workloads can be distributed, and what challenges are likely to arise as production increases. While no two projects are exactly the same, these experiences provide a foundation for making more informed decisions.

Our takeaway: Established procedures and previous operational experience help us approach new projects with greater awareness of the risks and resources involved. We believe that proper documentation, risk assessment, and continuous improvement are essential to maintaining reliable operations as projects grow in scale and complexity. These practices have also helped us develop a more systematic approach to managing multilingual data projects.

Final Thoughts: Large-Scale Data Projects Are Ultimately About People and Processes

After years of managing multilingual AI data and localization projects, one thing has become increasingly clear to us: production capacity is not determined by the number of contributors alone. It also depends on the systems supporting them, the instructions they receive, the quality standards they are expected to meet, and the people responsible for coordinating the entire operation. Even a highly capable workforce can struggle when the underlying processes are unclear or insufficiently organized.

Our experience has also taught us that scaling multilingual AI training data collection requires a balance between planning and flexibility. We need to establish clear requirements, realistic schedules, and appropriate quality controls while remaining prepared to respond to unexpected challenges. Just as importantly, we need to maintain honest communication with our clients and contributors, particularly when adjustments or improvements become necessary.

As artificial intelligence becomes more widely adopted, the demand for multilingual speech data, text corpora, and other forms of AI training and evaluation data is creating new opportunities for language service providers. We believe that experience developed through translation, localization, and large-scale linguistic operations can contribute meaningfully to this growing field, particularly in areas requiring human judgment, linguistic expertise, and careful coordination. At CCC, we continue to build on these experiences while refining our processes to support the development of technologies that can better understand and serve diverse linguistic communities.

About CCC’s Multilingual AI Data Services

Creative Connections & Commons Inc. (CCC) is a Philippines-based multilingual services company established in 2012. Building on our experience in translation, localization, and large-scale language projects, we provide multilingual AI data services, including speech data collection services, text corpus preparation, annotation, machine translation post-editing, linguistic evaluation, and quality assurance. Our teams support projects involving different languages, contributor profiles, production volumes, and technical requirements, with workflows adapted to the scope of each assignment.

Create A New Story.

Planning a large-scale multilingual AI data project? Let’s discuss how CCC’s experience in speech data collection, MTPE, and contributor management can support your production goals.