> For clean Markdown of any page, append `.md` to the page URL.
> For a complete documentation index, see https://docs.sarvam.ai/llms.txt.
> For full documentation content in one file, see https://docs.sarvam.ai/llms-full.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server.

# Speech to Text Best Practices

## Before You Transcribe

* Keep files within the supported formats (MP3, WAV, M4A, FLAC, OGG, AAC) and under 500 MB - if a source is larger, compress or trim it before uploading rather than fighting an upload timeout.
* Not sure what to expect? Try one of the built-in example clips first (e.g. "Gujarati finance news") to see how real code-mixed, domain-specific speech gets transcribed before you commit your own recording.
* Set the language to the actual spoken language, not a related or assumed one - this affects the model's phonetic expectations for the whole recording, not just word choice.
* Provide an accurate speaker count whenever you know it. For a two-person interview this is trivial; for a roundtable or panel, it's often the single biggest factor in transcript accuracy.
* If you're unsure of the exact count, round upward - merging an unused extra slot is a smaller error than splitting an undercounted one.

## Known Hard Cases

Recordings with cross-talk or overlapping speech (common in interviews and panels) will see diarization accuracy dip in exactly those overlapping segments - this is a known hard case for diarization generally, not specific to any one recording.

* Where possible, prefer recordings where speakers don't talk over each other.
* For recurring recording series with the same participants, note the correct speaker count once - you'll set it the same way every time, and getting it right the first time saves a re-transcription later.

## After Transcribing

* Review the transcript for consistency before exporting, especially around overlapping segments.
* Use the download or share icons on the Transcript panel to get a clean file rather than copying text directly from the screen.