message.comDevelopers
Developers/Concepts/Voicemail transcription

What is voicemail transcription?

Voicemail transcription converts a recorded voicemail into text that can be searched, filtered, triaged, and triggered on. Instead of listening to twenty voicemails in sequence, agents skim text and pick the urgent ones. message.com transcribes every voicemail automatically.

Definition

A voicemail is a piece of audio left on the platform when a caller is sent past the ring timeout. Transcription is the process of running automatic speech recognition (ASR) over that audio to produce a written record. The two outputs are stored together: the original audio file (for legal record and human verification) and the transcript text (for everything else).

How it works

We run a streaming speech-to-text pipeline:

  1. Caller leaves a voicemail. The carrier delivers the recording URL via the calls inbound webhook as a voicemail_left event.
  2. Server downloads the audio, hashes it, stores the file in object storage.
  3. Server sends the audio to a third-party transcription provider. The provider returns the transcript with word-level timestamps and confidence scores.
  4. The voicemail row gets the transcriptText, transcriptConfidence, and an array of word-level segments.
  5. An internal_note message is appended to the conversation containing the transcript so the agent sees it in the timeline.

Accuracy considerations

ASR is not perfect. Real-world voicemail audio is degraded by:

  • Mobile-network compression that strips frequencies above 3.4 kHz.
  • Background noise (cars, kids, dogs).
  • Accent variance.
  • Proper nouns the model has never seen.

message.com surfaces transcriptConfidence on every voicemail so agents know when to listen to the audio rather than trust the text. Confidence below 0.6 means the transcript is unreliable and the original recording should be played.

Transcripts make voicemails first-class members of the support data set. They are:

  • Searchable. Find all voicemails mentioning "refund" in the last week.
  • Routable. AI can classify intent and route the conversation to the right department before an agent looks at it.
  • Reportable. Voicemail volume and topic breakdowns show up in analytics.
  • Auto-answerable. The AI receptionist can draft a callback message based on the transcript.

API surface

Voicemails are exposed through the Voicemails API. Each row links back to its parent channel: "call" conversation. Transcripts also surface in the agent Socket.io namespace through the call:voicemail event.