Transcribe long recordings in your browser

A reliable first run, from preparing a recording to checking the exported transcript. Learn how model size affects completion, what happens locally and when another workflow may suit you better.

Open the transcription tool

Before you start

Use an up-to-date desktop browser. The browser and graphics driver determine which inference backends are available.
Make sure there is free memory: the browser needs space for the file, model weights and working buffers at the same time.
For the first run, choose a short clip with clear speech. This separates setup problems from issues caused by a long or noisy recording.
Plan to download the model first. Model and runtime downloads need internet access and may take time; later use depends on browser storage and cache.

Run the first job

Open the tool and choose Transcription. Under Step 1: Hardware Check, choose Test my processor. Cowslator uses browser capabilities and your device profile to recommend a model and processing path.
Under Step 2: Model Selection and Load, choose a recommended model. Start smaller on memory-limited devices and wait until model loading finishes.
Select one audio or video file, choose its spoken language or automatic detection, then press Transcribe. Use Translate to English only when you want translated English output.
Wait until processing finishes before exporting. The first run includes setup; compare later runs after the model is loaded.

For a folder, use the batch workflow and select its recordings. Files run in a queue to limit simultaneous memory use. Keep the tab open until the queue finishes, then check each result.

Review the result

Speech recognition is a draft. Replay short sections while reading and verify names, numbers, acronyms and specialized terms. Take extra care when speakers overlap, the microphone is distant, or background sound masks speech.

Choose TXT for plain text, SRT for subtitle editors and video players, or LRC for timed lines. Open the downloaded file in its destination application to confirm timing, encoding and line breaks.

What affects speed and stability?

Factor	What you may notice	Useful adjustment
Model size	Longer startup and higher memory use	Try a smaller recommended model on the same sample.
Browser backend	GPU support varies by browser, driver and device	Check the runtime status. A stable CPU/WASM run can be better than a failing GPU run.
Recording length	Long files keep buffers active longer	Test a short excerpt; split unusually long recordings into manageable sections.
Batch size	Large queues take longer and make failures harder to locate	Run a small group first and confirm its exports.

There is no universal speed claim because results depend on the full device and browser setup. To compare models fairly, use the same recording, language and device; wait for each model to load; repeat the run; and record model, backend, duration and completion.

Where files go

In single-device transcription, the browser processes media on your device and does not upload the recording to a Cowslator transcription server. Downloading the app and model still involves external software and model hosts. A collaborative session can send audio segments to paired devices on your local network.

Browser storage may retain settings and model data. Clear site data in browser settings to remove its local cache. The website also loads analytics and advertising services; see the privacy policy.