.png)
.png)
Jetlink Makes an Open Source Move: Conversations with JetlinkSTT Now Conversations are Data
- 1 day ago
- 4 min read
Consider this: A medium-sized call center handles 1,000 calls a day. Each call averages 5 minutes. That translates to approximately 83 hours of voice data per day.
Within those 83 hours, there are real customer issues, recurring complaints, moments that either convert or cancel purchases. But the vast majority of this data is never analyzed. Because the voice doesn't fit into the tables.
This is where speech recognition technology — Speech-to-Text (STT) or Automatic Speech Recognition (ASR) — comes into play. And at Jetlink, we've developed our own model in this area and launched it: JetlinkSTT-v1.0-7B, now available as open source on Hugging Face.
What is JetlinkSTT and what is it not?
JetlinkSTT is not a transcription service. It's an AI model that you can integrate into your own infrastructure and customize to your needs.
Built on the Qwen3-ASR architecture, the model is designed to operate in both research environments and enterprise production pipelines. With 2 billion parameters and BF16 accuracy, it is both lightweight enough and powerful enough to deliver reliable performance in real-world conditions.
What does it do, in short?
It records the audio and converts it to text.
It automatically identifies the spoken language (Language Identification).
It operates in both real-time (streaming) and batch (offline) processing modes.
Long recordings — including hour-long meetings — work seamlessly.
It offers strong accuracy even in noisy environments, with accented speech, and even in audio files with background music.
52 Language Support: Turkish is also on the list.
For global businesses, language support isn't a feature—it's a necessity. The Qwen3-ASR family, on which JetlinkSTT is based, encompasses 52 languages and dialects . These include Turkish, English, German, French, Arabic, Japanese, Korean, Russian, and many other commonly used business languages.
Given the scarcity of sufficiently mature and tested models for Turkish voice data processing, this feature alone represents a significant differentiator. It offers real use value for both local businesses serving the Turkish market and global brands managing a multilingual customer base.
Who can use it and in which scenarios?
Call Centers and Customer Service Teams
In call centers, voice data is the rawest and most honest source of customer experience information. But this data often remains limited to a small sample randomly listened to by quality control teams.
With JetlinkSTT, all conversations can be automatically transcribed, categorized, and analyzed. Which topics cause the most problems? Which agents provide solutions faster? At what points do customers disconnect? The answers to these questions lie within the voice data—you just need to extract it.
Meeting and Conversation Automation
Sales meetings, internal meetings, customer onboarding sessions — they're all recorded, but how many of them are actually processed?
JetlinkSTT is designed to successfully handle long audio files. You can instantly obtain and summarize a transcript of an hour-long meeting, extract actions, and integrate them into pipelines that differentiate between different speakers.
Subtitling and Content Production
Video content, educational materials, podcasts — subtitling and transcription are a significant operational burden for all of these. By connecting the model to subtitling production pipelines, you can completely automate this process.
Voice Assistant Infrastructure
When building a voice assistant system, the most critical layer is the accurate conversion of speech to text. Understanding accuracy begins here. JetlinkSTT delivers the performance needed to build this layer on a solid foundation — both for real-time response in streaming mode and for batch processing.
Speech Analytics and Research
For those who want to perform sentiment analysis, topic classification, or extract behavioral models from voice data, JetlinkSTT forms the first step in the pipeline. It's impossible to build the analytics layer without first converting raw voice into structured data.
How does it work technically?
JetlinkSTT is packaged to integrate directly with the Transformers library. Additionally, thanks to vLLM support, it can be effectively used in production environments requiring higher throughput and lower latency.
Hardware requirements vary depending on the use case:
8 GB of VRAM is sufficient for development and testing environments.
For a more comfortable experience, 12–16 GB of VRAM is recommended.
Datacenter-class GPUs are ideal for high concurrency and production loads.
Installation is extremely simple:
conda create -n jetlinkstt python=3.12 -yconda activate jetlinksttpip install -U qwen-asrThe model was released under the Apache 2.0 license — fully open for commercial use, modifiable, and distributable.
Why did we release it as open source?
Developing products in the field of artificial intelligence and contributing to the ecosystem are two different approaches. At Jetlink, we choose the second path.
For developers: You can fine-tune, quantize, and integrate with different serving stacks according to your own use case. You have complete control over the model.
For researchers: A benchmark that can be used in studies on organizational speech recognition, and a point of reference for comparative analyses.
For enterprise teams: You can host it on your own infrastructure and deploy it in a way that suits your data privacy requirements. Fully compatible with on-premise operations, without cloud dependency.
This approach aligns with Jetlink’s larger vision: technology that transforms customer interactions should not be found in closed boxes—it should be built on transparent, testable, and scalable foundations.
Connecting with JetVoice: From Model to Product
JetlinkSTT is not just a standalone model — it forms one of the technological foundations of Jetlink’s voice-based AI product, JetVoice .
JetVoice delivers real-time voice experiences: human-indistinguishable responses in customer service, automated workflows, and voice-command triggered processes. JetlinkSTT provides the fidelity layer for this experience. If you want to integrate the model into your own infrastructure, you can access it directly via Hugging Face. If you're looking for a ready-to-use, enterprise-level solution, JetVoice is exactly what you need.
Access the Model
JetlinkSTT-v1.0-7B is now in open access on Hugging Face:
You can speak with our team to change how you process voice data, integrate AI into your call center operations, or explore how you can use JetVoice in your organization.
Comments