Explore MERaLiON 3 10B

Uploads the audio recording for transcription and spoken-language understanding. Clear speech and limited background noise improve results; long recordings may need chunking and overlap settings to control memory without cutting words at boundaries.

Controls timestamp detail in the transcript. False returns text only, True returns segment-level timing, and Word requests timing for individual words; finer timestamps increase response detail and may be less precise on noisy audio.

Splits long audio into chunks of this many seconds before transcription. Chunking controls memory use and enables long recordings, but chunks that are too short provide less linguistic context and may reduce accuracy.

Adds overlapping audio on the left and right of each chunk, supplied as one number or a [left, right] pair in seconds. Overlap reduces cut-off words at boundaries but increases duplicated processing time.

Controls how many items are evaluated in parallel during one forward pass. Larger batches can improve throughput but use more memory; reduce the value if processing runs out of device memory.

Provides advanced keyword arguments directly to the model’s generation method. Use it for supported controls that are not shown separately in the form, and avoid duplicating named settings because these values may override or conflict with them.

Suppresses the pipeline warning shown when external chunking is used with Whisper’s own long-form behavior. It changes only warning output, not transcription; enable it only after confirming the chosen chunking approach is intentional.

Passes a JSON object of keyword arguments to the pipeline decoder during transcript post-processing. Use it for decoder-specific controls not exposed elsewhere, and leave it empty for the model defaults because unsupported keys can cause decoding errors.