Voice
Sheet 11/12 · Medium–High · 4–12 weeks (pilot first)Calls go unanswered at peak, the IVR is a maze, and customers send voice notes nobody transcribes.
- The landline rings out during lunch and after 17:00; the mobile is engaged.
- Customers send WhatsApp voice notes in Shona or Ndebele; a person must listen to each one.
- The IVR has seven options and callers press zero.
- Appointment confirmations and payment reminders are made by hand, one call at a time.
- Call outcomes are not recorded, so nobody knows the top reasons people call.
- Answer a phone call in English, greet, understand the reason, book or confirm an appointment, take a message, and hand to a person.
- Transcribe and summarise WhatsApp voice notes so staff read instead of listen — with mixed-language support that must be tested on your customers.
- Make outbound confirmation and reminder calls or messages on a schedule.
- Log every call with reason, outcome and a transcript.
- Handle Shona and Ndebele speech at the quality it handles English — recognition for these languages is still maturing (academic work and "coming soon" commercial tools as of 2026).
- Survive poor audio: a kombi in the background, a weak signal, a speakerphone.
- Take a payment over the phone without a separate, compliant payment step.
- Replace the receptionist for anything emotional, urgent or unusual — it must hand over fast.
| Effort band | Medium–High · 4–12 weeks, pilot first |
|---|---|
| Data you need | Call reasons and volumes by hour; booking rules; recorded calls with consent; scripts |
| Integration points | Telephony (mobile/SIP), WhatsApp voice notes, bookings/CRM |
| People | A receptionist or agent who takes hand-offs; a language reviewer |
| Measure | Answer rate; calls resolved without a person; hand-off time; caller satisfaction; language error rate |
| Zimbabwe constraint | Sixteen official languages, Shona and Ndebele dominant in speech; voice-note culture on WhatsApp; audio quality on mobile networks |
| Code | Shape | Fit | Why |
|---|---|---|---|
| S-02 | Pilot | Primary | Two to four weeks on one number with real callers and a person listening to every call; the language question is answered here, nowhere else. |
| S-03 | Implementation | Secondary | Call flows, booking logic, hand-off rules and voice-note handling built after the pilot's evidence. |
| S-04 | Integration | Secondary | Bookings, CRM and the telephony provider (SIP trunk, mobile number) are the connections that make it useful. |
| S-05 | Training | Secondary | Front-line staff learn when to trust a transcript and when to listen. |
| S-06 | Managed support | Later | Voice models and telephony change fast; monitor call quality monthly. |
- Play us three of our own recorded calls (with consent) through the system: one English, one Shona, one mixed with background noise. What did it get wrong?
- What is the word error rate you measured on Shona or Ndebele, on whose recordings, and how recently?
- How does the caller reach a person — how many seconds, how many attempts — and what does the hand-off sound like?
- Which telephony does this run on (a Zimbabwe mobile number, SIP, WhatsApp calling), and who owns the number?
- Where is call audio processed and stored, in which country, and for how long?
- What does a call cost at 1,000 and 10,000 minutes a month, including telephony and the speech model?
The full bank of 60 questions, grouped by risk, is at /vendor-questions/.
Language coverage in 2026, honestly
Zimbabwe’s constitution recognises sixteen official languages; Shona is spoken by the large majority and Ndebele by a substantial minority, and everyday speech mixes them with English. Commercial English voice agents are mature. For Shona, the picture is: research systems exist (a 2025 study built a Shona-slang dataset and an intent classifier reaching 96.4% accuracy on text, and separate work has trained deep-learning speech recognisers for Shona); some international speech-to-text providers list Shona; and at least one African language-tooling company lists Shona speech recognition and text-to-speech as “coming soon”. Ndebele has less published work. Treat any vendor’s Shona voice claim as a pilot question, not a purchase.
| Task | English | Shona | Ndebele |
|---|---|---|---|
| Inbound phone agent | Ready to pilot | Pilot with a person listening | Pilot with a person listening |
| Voice-note transcription (WhatsApp) | Ready | Usable for gist; verify names, amounts | Usable for gist; verify names, amounts |
| Outbound reminders (text-to-speech) | Ready | Test pronunciation with staff first | Test pronunciation with staff first |
| Intent from short text (Shona/English mix) | Ready | Good on text (research results) | Less evidence |
Why “transcribe the voice note” is the first win
The cheapest voice project is not a phone agent: it is turning the forty daily WhatsApp voice notes into text summaries so a person reads them in one minute instead of listening for twenty. Accuracy needs to be “good enough to know what it is about”, names and amounts get verified by ear, and nothing faces the customer. Run that for a month before you buy a phone agent.
The hand-off test
Call the vendor’s demo number and say something unexpected in your normal mix of languages. Count the seconds until a person or a graceful message. Anything that loops through a menu twice fails.