Voice transcript routing
Early access. Early access Voice and phone support is built but not switched on for every account yet. Email [email protected] to join early access. This page describes how it works once it is on for your workspace.
How transcription works
During and after each call:
- Twilio streams audio to AskVault.
- ASR (automatic speech recognition) transcribes in real time.
- Bot analyzes for intent, sentiment, named entities.
- Routing rules apply post-call: assign team, tag, escalate.
- Transcript available in the conversation view within 60 seconds.
Accuracy is best on clear audio and drops with background noise and strong accents. We have not published a benchmark.
Routing patterns
Three common patterns:
By topic. Transcript mentions "refund", route to billing team. Mentions "down" or "outage", route to engineering.
By sentiment. Negative sentiment detected, assign senior agent immediately. Positive sentiment plus question resolved, auto-close.
By caller identity. Verified Enterprise Plan customer, route to dedicated AE. Anonymous or Free Trial caller, route to general support.
Configure under Automation Rules using transcript content as the condition.
Sample routing rule
For an SaaS:
- Trigger: voice call ended.
- Conditions: transcript contains "cancel" OR "refund".
- Actions: tag
billing, assign to billing-pool, notify #billing-alerts in Slack.
Apply rules per call. Routing typically completes within 5 minutes of call end.
Transcript structure
Per call:
- Full transcript with speaker labels (Caller / Bot / Agent).
- Timestamps per utterance.
- Confidence scores per word.
- Detected entities (names, dates, order numbers).
- Sentiment trajectory across the call.
Visible under Conversations > [call] > Transcript.
Searching across transcripts
Like text conversations, voice transcripts are full-text searchable:
- Conversations > Search.
- Filter channel = voice.
- Type the search query.
- Results show matching call segments with timestamp.
Useful for QA, training, and pattern analysis.
Multi-language transcription
Auto-detection plus explicit selection:
- Auto-detect from first utterance. Switches mid-call if caller switches.
- Workspace default language if auto-detect fails.
- Several languages supported. Accuracy varies by language.
Accuracy varies by language and audio quality, and is usually highest for English. We have not published per-language figures.
Speaker diarization
For multi-party calls (caller + bot + transferred-agent):
- Each speaker labeled distinctly.
- Transcript shows who said what with timestamps.
- Audit-ready for compliance.
Accuracy is lower on calls with three or more parties. We have not published a figure.
Privacy and PII redaction
For sensitive content:
- Automatic redaction. Email addresses, social security numbers, card numbers and long runs of digits are masked in the call summary and transcript AskVault stores.
- No setting to turn it on or off. It is always applied to what AskVault stores.
Redaction is pattern-based, so it can miss details written in an unusual format. Do not treat it as a compliance guarantee.
Call recording retention
- Audio recordings are separate from transcripts and follow your Twilio storage settings.
- Transcripts kept for 90 days on every plan, then deleted.
Useful for legal hold or extended training.
Integration with downstream tools
Send transcripts to other systems:
- Webhook on call end. Push full transcript to your CRM.
- HubSpot Activity log. Transcript attached to contact record.
Configure under automation rules with webhook actions.
Analytics
Per workspace:
- Calls per day.
- Average call duration.
- Bot-resolved vs human-resolved split.
- Most common topics (from transcript clustering).
Visible under Analytics > Channel: Voice.
Limits
- Transcription latency. Real-time during call; final transcript within 60 seconds post-call.
- Max call duration. 30 minutes; the call then ends.
- Recording size. Variable; about 1 MB per minute of audio.
- Routing rules per workspace. 100 (Business and above). Voice is a Business-plan feature; Growth and below don't have call routing rules.
Common pitfalls
Transcription quality low. Background noise or poor audio. Improve caller audio (request hands-free or quiet environment).
Wrong routing. Conditions too broad. Tighten with multi-word phrases.
A detail was not redacted. Redaction is pattern-based and can miss unusual formats. Contact [email protected] with the example.
Transcript missing words. ASR limit. Cross-check with audio recording.
FAQ
Can I export transcripts as CSV?
Yes via API or Conversation Export. Includes timestamps and confidence scores.
Does transcription work for voicemail?
Yes. Voicemail audio transcribes within 60 seconds.
Is the transcript searchable in the customer's conversation history?
Yes. Cross-channel search treats voice transcripts like any other message.