
Weekend Showcase Challenge: Bêje
Dengbêj AI continues its journey as Bêje gains a Kurmanji voice. This showcase explores the missing Kurdish speech tools, the KurdishTTS integration, and how AWS helped transform a storytelling experiment into an autonomous bilingual news radio.
Dengbêj AI has grown through several AWS Builder challenges. It began as an experiment in reimagining Kurdish storytelling with generative AI. Later, it became Bêje, an always-on Kurdish news-radio agent that prepares new programs before visitors open the application.
For this summer showcase, I wanted to add one important feature that had been missing from the beginning: a Kurdish voice.
From Dengbêj AI to Bêje
The word dengbêj joins deng—voice—and bêj—to tell or recite. Voice is not an optional detail in the Dengbêj tradition. It is at the center of it.
The first version explored how AI could transform written material into stories inspired by Kurdish oral narration. The next version developed into Bêje, an automated news-radio experience with daily and regional programs.
Bêje now collects recent stories, selects and classifies them, generates Kurdish headlines and summaries, creates broadcast scripts, and publishes the results through a bilingual interface. Visitors can choose programs such as Today, World, Middle East, Turkey, Kurdistan, or Rojava.
However, one contradiction remained: the application could write Kurdish, but it could not speak Kurdish.
The Missing Voice
The existing pipeline was already producing complete Kurmanji scripts using the Latin alphabet, including characters such as ê, î, û, ç, and ş. However, Amazon Polly does not currently provide a Kurdish voice. The original radio therefore used English narration through Polly.
I investigated other major speech providers and models. Amazon Nova Sonic did not list Kurdish support, and the Qwen3-TTS model available through SageMaker JumpStart did not include Kurdish among its supported languages. I also checked Google Cloud Text-to-Speech and Azure Speech, but neither offered the Kurdish voice required for this project.

My investigation showed a recurring problem: the Kurmanji scripts were ready, but Kurdish voices were missing from the major speech platforms I evaluated.
This is more than a technical inconvenience. When a language is missing from foundational tools such as text-to-speech, developers cannot offer the same accessibility and experience to every community. It affects storytelling, education, news, assistive technology, and the ability of a language to participate fully in new digital products.
The text was there. The missing tool was the voice.
Giving Bêje a Kurmanji Voice
I eventually found KurdishTTS, a specialized service supporting both Kurmanji and Sorani. I integrated its API into the existing AWS pipeline while keeping Amazon Polly for English narration.
Bêje can now select audio according to the interface language. Kurdish mode uses the new Kurmanji recording when one is available, while English mode continues using Amazon Polly. If a Kurdish recording is missing, the player safely falls back to English.
The Kurdish narration is currently an experimental beta feature. I deliberately introduced it through a separate development branch, tested it, and then merged it into the current Bêje version without modifying the frozen versions used for earlier challenges.
How I Built It
The application uses a serverless AWS architecture:
- Amazon EventBridge starts the news and program workflows on a schedule.
- AWS Lambda ingests, processes, classifies, and publishes stories and audio.
- Amazon Bedrock generates Kurdish headlines, summaries, and radio scripts.
- Amazon Polly provides English narration.
- Amazon DynamoDB stores stories, programs, processing state, and TTS quota usage.
- Amazon S3 stores the generated English and Kurdish audio.
- AWS Secrets Manager protects the KurdishTTS API key.
- AWS Amplify hosts the public application.
Integrating an external speech provider also introduced a new responsibility. Its free API allowance is limited, so I added sentence-aware text chunking, persistent monthly quota tracking, atomic character reservations, failure refunds, feature flags, and a safety limit.
Scheduled Kurdish synthesis remains disabled while I evaluate quality and usage. This protects the free allowance while allowing the current feature to be demonstrated. The project now has 255 passing tests covering the application and its audio pipeline.
What I Learned Across the Summer
The AWS Builder challenges have kept me motivated to return to Dengbêj AI and improve one meaningful part at a time. Instead of trying to build everything in one step, I learned how to freeze stable releases, experiment on separate branches, protect secrets and quotas, test failure paths, and treat deployment as part of development.
The challenges have already rewarded me with three jackets and AWS credits, but the most valuable reward has been the experience. Each challenge pushed the project into a new stage: first an idea, then a working creative application, then an autonomous agent, and now a more complete bilingual radio experience.
I also want to recognize the wider AWS Builder community. Seeing other builders continue polishing and sharing their projects throughout the summer encouraged me to keep developing mine.
What Comes Next
The next steps are to compress the Kurdish audio files, evaluate voices with more Kurdish listeners, improve coverage across regional programs, and explore Sorani narration.
Dengbêj AI is still growing, but this showcase version has reached an important point. Bêje no longer only displays Kurdish words—it can now speak them.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article