How a GCC Telco Built an Arabic Speech-to-Text Dataset from Call Archives

A GCC telecom operator transformed 10,000 archived customer calls into a high-quality Arabic speech-to-text training dataset using Munsit STT and expert annotation. The resulting Gulf dialect dataset improved intent-classification accuracy and created a scalable foundation for future AI model development.

The Challenge

Telcos building AI for customer-facing applications, intent classification, sentiment analysis, churn prediction, and virtual agents need training data. For Arabic-language models, that data is hard to find. Publicly available Arabic speech datasets are mostly Modern Standard Arabic sourced from news broadcasts. They don't reflect how customers actually speak when calling a telco in the Gulf.

A GCC telco's data science team had a clear use case: fine-tune an intent classification model covering billing, technical support, plan changes, complaints, and service requests. The training data had to reflect real customer language, Gulf dialect, code-switching, and the specific product vocabulary their customers used every day.

They already had hundreds of thousands of call recordings in their archive. In theory, this was exactly the training source they needed. The recordings were unusable for training without transcription, diarization, and labeling.

The Data Pipeline Problem

Transcribing and labeling call recordings at scale is expensive and slow. The team had received a quote from a general transcription provider for 10,000 calls. The cost was high, the turnaround was weeks, and the provider had no specific capability in Gulf Arabic, meaning every transcript would need native speaker review before it could be used for training.

What the team needed was a pipeline that could handle three things:

• Transcribe Arabic calls at scale with Gulf dialect accuracy
• Segment transcripts by speaker so customer utterances could be extracted separately from agent responses
• Pass the resulting text to a labeling workflow where customer intent could be classified

The Approach

CNTXT AI addressed both stages of the pipeline directly. Munsit STT processed a batch of 10,000 call recordings from the telco's archive via the API in batch mode. Each call was returned as a speaker-diarized transcript, with customer utterances automatically extracted and separated from agent turns.

Those customer utterances were then passed to CNTXT AI's Arabic data annotation team for intent labeling. Annotators classified each utterance against a 28-category taxonomy built jointly with the telco's data science team, covering intent categories specific to their service context, not generic call centre categories. Quality control included double annotation on 15% of utterances, with inter-annotator agreement measured and resolved by a senior reviewer.

The final output was a labeled Arabic speech dataset specific to the telco's customer interaction domain and formatted for direct use in the team's fine-tuning workflow.

Results

The data science team had a usable training dataset within six weeks of project start. The manual route would have taken months.

The intent classification model fine-tuned on this dataset outperformed the version trained on public Arabic data on the telco's internal evaluation set. The improvement was most visible in two areas:

• Gulf dialect inputs, where public MSA training data consistently fell short
• Product-specific terminology, vocabulary that appeared frequently in the telco's calls but was absent from broadcast Arabic datasets

Beyond the initial model improvement, the team now has a repeatable pipeline for expanding training data as the product portfolio grows and new intent categories emerge. That same pipeline is currently being used to build a sentiment analysis training set from a separate call sample.