In August 2026, we began transcribing audio and extracting text from documents on our own hardware. Until then, the document-reading workflow we examined relied on an external provider. Over the following weeks, the local machine took on roughly four out of every five pages in that workflow.
We started with specific tasks: turning speech into text that stays aligned with its recording, and reading a page without losing names, numbers or references. We tested open models on legislative material, brought them together on one machine and kept an external option for work it could not finish.
Testing on the material we use
A speech can include names, interruptions and silence. A document can mix paragraphs and tables, or arrive as a scanned image. These differences matter when choosing a model: recovering many words is not enough if a number changes or the relationship between two columns disappears.
Our audio tests included about 170 segments across seven language groups, with coverage varying by model. The document tests covered 60 pages from ten official sources. We compared the output with reference transcripts and text embedded in PDFs to identify missing or changed words.
From source material to text
Returning to the moment someone spoke
Whisper large v3 turbo converts audio to text through automatic speech recognition, or ASR. It also records when each word starts and ends, so we can return to that point in the recording.
Its language coverage was another reason we chose it. In the configurations we tested, Granite NAR repeated text during silence, and Parakeet introduced English words into a Spanish passage.
Preserving what a page says
When text is part of an image, OvisOCR2 extracts it through optical character recognition, or OCR. The aim is to recover both the content and the relationships that make it readable: which number belongs to a row, where a section begins or how a sentence continues.
Across the 59 pages we evaluated, Ovis came close to the external service when comparing recovered words. We chose it over PaddleOCR, which used less memory, because it preserved more of the text in that test. We excluded one cover page whose PDF did not contain a complete text reference.
Selecting what needs further analysis
We later added Qwen3.5 9B to compare the subject of a hearing with the topics we follow. For example, a hearing about water supply may be relevant to someone following water resources. The large language model, or LLM, makes this initial selection.
The system then analyzes the content of the selected hearings to prepare alerts. We added Qwen after the period covered by our analysis.
Three tasks on one machine
The initial tests ran on a machine with 8 GB of video memory. The current machine has 16 GB.
Available memory shapes how we divide the work. Audio is processed in segments and documents one page at a time. Models stay loaded between jobs to avoid repeated startup. Qwen runs through llama.cpp and uses both GPU and system memory. This brings the three tasks together without requiring every model to fit entirely on the GPU.
The shift to local processing
From August 1 to 16, the external provider processed every page recorded in the OCR workflow, about 26,000 in total. From August 17 to October 5, our server processed over 50,000 pages, around 80% of the total for that period. The remaining pages continued to use the external service.
Source: aggregate OCR records. Each bar represents 100% of the pages in its period. Cutoff: October 6, 2026, 00:00 UTC.
During that second period, the server also completed about 33,000 audio jobs, totaling roughly 2,700 hours of processed segments. These are accumulated segment durations, not unique session hours. Another 390 audio jobs were routed to the fallback provider.
Since the periods differ in length and material, we compare the share of pages processed on our own hardware and by the external provider.
Keeping another route available
Moving processing onto our own server also means handling interruptions. If the machine is unavailable or a job cannot finish after retries, the system can use the external provider. Both routes return the result to the same workflow.
We also adjusted how the system handles delays and cancellations. A slow download is still recognized as active work. Cancelling a document stops the remaining pages from being read.
Keeping both routes available lets us introduce a model gradually: test it on our sources first, then decide which work it will handle.
When we started, we wanted to see how much of our work we could handle on our own hardware. Today, that machine is part of our daily operations. That experience helps us decide what else we can build ourselves.