Most transcription wrappers accept whatever you hand them and fail somewhere in the middle when the sample rate is wrong. This one names PCM16, 16 kHz, mono up front and ships the conversion line next to the requirement, which turns an opaque failure into a preprocessing step. stt_health and stt_pricing exist so the failures that have nothing to do with your audio are checkable in advance: whether the service is running, and what a minute costs — the pricing call charges nothing to ask.
A speech-to-text server, exposing 3 tools that transcribe a local WAV file through dTelecom and report the service's pricing and health.
- Transcription from a path on disk: transcribe_file takes a file path, a language and a minutes argument, so the billable quantity is visible at the call site.
- One accepted input shape, stated exactly: the file must be PCM16, 16 kHz, mono — and the tool description carries the ffmpeg invocation that converts anything else into it.
- Price checked before commitment: stt_pricing returns current pricing and explicitly requires no payment to call.
- A liveness check that is separate from a failed transcription: stt_health reports whether the service is running.
A dTelecom private key, supplied as DTELECOM_PRIVATE_KEY, and audio already in PCM16 16 kHz mono WAV at a path the server can read.
One command plus a key — npm install -g @dtelecom/stt-mcp, then supply credentials