Aramb Voice

Aramb Voice Toolkit

Two tools turn audio into text and text into audio, so an agent can read the voice notes people send and reply with a spoken voice note.

• aramb_mcp.voice_transcribe — an incoming audio file → its text.
• aramb_mcp.voice_synthesize — your text → an audio file, returned as a blob artifact you deliver through aramb_chat.

When to use this

• A user sends a voice message and you need its contents → voice_transcribe.
• You want to answer out loud (a spoken reply on a chat that supports voice notes) → voice_synthesize, then deliver the result.
• Accessibility or hands-free flows where a spoken reply is expected.

Do NOT use this to place a real phone call (that is aramb_phone), or to send an ordinary text message (that is aramb_chat). Synthesizing audio is only worth it when the recipient actually wants a voice note.

CRITICAL: mcporter syntax rules

• ALL arguments MUST use key="value" format (NOT positional args).
• Do NOT use --output — it is not supported by mcporter call.

Tool 1 — voice_transcribe (voice note → text)

When a voice note arrives it is attached to your message like any other file, and shown to you as a path, e.g. @/…/attachments/<id>-note.ogg. Pass that path exactly as shown — the platform resolves it to the stored audio for you; you never handle a storage key yourself. A leading @ is optional and a bare absolute path also works.

Args:

• audio (required) — the attachment path of the voice note, as shown to you.
• language (optional) — a hint like en; omit to auto-detect.

npx mcporter call aramb_mcp.voice_transcribe audio="@/…/attachments/<id>-note.ogg"

Returns:

{ "text": "the spoken words, transcribed", "language": "en", "duration_seconds": 2.9 }

Read text and continue as you would with any user message.

Tool 2 — voice_synthesize (text → voice note) + delivery

voice_synthesize speaks your text and returns a blob artifact descriptor — it does not send anything on its own. To get the voice note to the user you pass that descriptor to aramb_mcp.chat_deliver_artifacts (see the aramb-chat skill).

Args:

• text (required) — what to say. Keep it to what should be spoken aloud (no markdown, links, or code — those do not read well as speech).
• format (optional) — ogg (default) or mp3. Prefer ogg: it is the voice-note format chat apps expect, so it plays inline as a voice message.
• voice_id (optional) — a specific voice from the voice catalog; omit for the default voice.

npx mcporter call aramb_mcp.voice_synthesize text="Your order ships tomorrow — talk soon!" format="ogg"

Returns a ready-to-deliver blob:

{ "kind": "blob", "s3_key": "…", "content_type": "audio/ogg",
  "filename": "voice-note.ogg", "size": 4841 }

Delivering it

Pass it to deliver_artifacts in the artifacts array. The only required field is s3_key; the platform infers the audio type from the key, so a minimal delivery is just kind + s3_key. Optionally add name (display filename) and mime_hint (e.g. audio/ogg):

npx mcporter call aramb_mcp.chat_deliver_artifacts \
  artifacts='[{"kind":"blob","s3_key":"…","name":"voice-note.ogg","mime_hint":"audio/ogg"}]'

The delivery layer re-validates the s3_key against your session and generates a fresh download/playback URL when the message is sent, so the recipient gets a playable voice note. You never construct or presign a URL yourself.

Field-name note: voice_synthesize reports the audio as content_type / filename, but deliver_artifacts reads the equivalent optionals as mime_hint / name. Either way s3_key is the field that matters — the rest is inferred if omitted.

End-to-end pattern (reply to a voice note with a voice note)

1. A voice note arrives → call voice_transcribe with its @path → read text.
2. Decide your reply and keep it speakable (plain sentences).
3. voice_synthesize your reply (format="ogg") → get the blob descriptor.
4. deliver_artifacts with that descriptor → the user hears your answer.

Notes

• Supported voice-note formats include ogg/opus, mpeg (mp3), wav, mp4/m4a, webm, and opus. If a sender's clip is in another container, transcription may return empty text — ask them to resend as a standard voice note.
• Synthesis and transcription each run a real speech model; use them when a voice interaction is actually intended, not for routine text turns.