Making Armenian heard.
Exploring how speech models can serve a language and community I’m already part of.
Armenian
The language I’m focused on
Fine-tuning + evaluation
Training work and transcript comparisons
Live transcription
The application I’m working toward
Why Armenian speech
Armenian is my native language, and I spend a lot of time around Armenian speech and music through church. Speech recognition is especially interesting to me because the use case is right in front of me: someone speaking, and useful Armenian text coming out.
The longer-term goal is low-latency transcription of spoken sermons. A system for that setting has to handle the speaker and the language well, while producing text quickly enough to be useful.
What I have worked on
I have fine-tuned Whisper on Armenian Common Voice data and worked with models such as Whisper large-v3-turbo. I have also been experimenting with Qwen ASR and looking at how to adapt it for Armenian.
The work includes comparing transcripts and thinking carefully about the data and evaluation. A transcript can look plausible while still changing a word or phrase that matters.
Accuracy and latency are connected
A model that produces a good transcript after a long wait solves a different problem from a live transcription system. I’m interested in both the quality of the text and the practical cost of getting it while someone is still speaking.
That is why I also spend time on local inference, GPU memory limits, and the hardware the system might actually run on. I want the eventual application to make sense outside a training notebook.
What I’m still exploring
I’m exploring low-resource fine-tuning, ways to use available Armenian audio more effectively, and how to evaluate live transcription. Qwen adaptation and streaming are areas of ongoing investigation.
This is an active project. My goal is to build toward a useful transcription tool, and I keep the training experiments separate from claims about a finished live system.