Your customers answer in Sundanese. Now your reminder calls do too.
Morfem calls customers before a payment is due, in Indonesian, Sundanese or Javanese. It costs about a third of a call center agent, and every amount it reads out comes from your system, not from the AI.
Illustrative spectrogram of the call line. Turmeric marks data read from the client's system, never made up by the AI.
Same sentence, same data, three languages. 30-day pilot for 500 customers. No fee, no contract.
From customer list to report in four steps
- Step 1
You send the customer list
By API or a daily CSV file: name, phone number, amount, due date and preferred language.
- Step 2
Morfem calls during permitted hours
The assistant opens in Indonesian, then switches to Sundanese or Javanese if the customer answers in that language.
- Step 3
Amounts are read from your data
The language model only picks the sentence. Amounts, dates and names come from your system.
- Step 4
Results land in your dashboard
Promise to pay, asked for a person, wrong number or no answer. Every call has a recording and a transcript.
Follow one call from start to finish
Connected to your systems and knowledge base
Morfem takes customer data from your systems, calls within the rules, then sends the results back. The language model never writes a number itself.
Core banking / LOSin the demo
name, number, instalment, due date
CRMvia webhook
Qontak, HubSpot, Salesforce, Zoho
Hospital systemin the demo
clinic schedule, patient record
Knowledge basecoming soon
FAQ, how to pay, branch address, SOP
data from your system
Morfem
- Locked dataamounts and dates filled from your system, not the model
- OJK rules08:00 to 20:00, identity first, audit log
- Three-language voiceIndonesian, polite Sundanese, polite Javanese
call · result webhook
Customer
takes the call, asks, promises to pay
Your dashboard and CRM
promise to pay, asked for a person, recording, transcript
Our own speech model, tested on speakers it never heard
We trained Javanese and Sundanese speech recognition on public data. Test speakers and sentences were not used in training.
Javanese
Phone recordings (OpenSLR 35, 400 sentences)
Studio recordings (OpenSLR 41, 300 sentences)
Free sentences (OpenSLR 35, 262 sentences)
Sundanese
Free sentences (OpenSLR 36, 153 sentences)
Character error rate (CER) after training: Javanese 5.7% on phone recordings and 7.8% on studio recordings, Sundanese 7.7% on free sentences. Some word errors differ by one or two letters only, such as é written as e.
All numbers come from read sentences, not real customer calls. We measure real calls during the pilot. Free sentences do not follow the sentence patterns that repeat in the training text. Through a simulated phone line, WER on free sentences is 39.1% for Javanese and 42.8% for Sundanese. A third of the Javanese phone-recording test sentences follow sentence patterns that also occur in the training text.
Listen for yourself
Recordings from the Javanese test set (OpenSLR 41). Neither the sentences nor the speakers were used in training.
- What was said
- urutan kaping rong puluh yaiku sekolahku
- Whisper, untrained
ឍ- Morfem model
- urutan kaping rong puluh yaiku sekolahku
- What was said
- sampo ning rudy hadisuwarno salon sak paket karo pelembape
- Whisper, untrained
samphounṅkhrudiḥaṭiṣuārnoṣalónṣaṅkhketṅkhkaraṅkhkaraṅkhkaraṅkhkaraṅkhkaraṅkhk…- Morfem model
- sampo ning
rudihadiswarnasalon sak paket karopelembapke
- What was said
- lipene the face shop nyolok nyolok kabeh wernane
- Whisper, untrained
līpnēdēufēishèpnǐòlònǐòlòkāpēhōrnāmī- Morfem model
lippinethe face shop nyolok nyolok kabeh wernane
- What was said
- emma roberts njaitna klambi apik gawe buwuh nang rabine kancane
- Whisper, untrained
emarobertjaitnoklambyapeekhawebuwohnangrappinikonconi- Morfem model
- emma roberts njaitna klambi apik gawe buwuh nang rabine kancane
- What was said
- nggih ndang dipunasahi piringipun mung namung mboten wonten tigang menit
- Whisper, untrained
ghehindangdipunasahipiringipunmengunamengbotenwontantigaangmenit- Morfem model
- nggih ndang dipunasahi piringipun mung namung mboten wonten tigang menit
Struck-through words differ from what was said.
Audio: Google, OpenSLR 41 and 44 (Kjartansson et al., 2018; Sodimana et al., 2018), CC BY-SA 4.0 licence.
For lenders whose customers live in West, Central and East Java
The 2010 census counted 68 million Javanese speakers and 32 million Sundanese speakers. Almost all of them understand Indonesian, but many rural bank customers and older patients respond faster when the call is in the language they use at home.
Rural banks (BPR and BPRS)
Instalment reminders 3 days before the due date, in the language customers use at the market and at home.
Finance companies
Reminders for motorbike and electronics loans to thousands of customers a day, with each promise to pay on record.
Cooperatives, hospitals and clinicsComing soon
Cooperative instalment reminders and follow-up appointment confirmations. We open these after the lender pilots finish.
Work out your monthly reminder cost
Change the numbers on the left. The comparison uses cost per talk minute, because that is what you pay a call center for.
15,000 talk minutes per month
- Call center agents
- Rp12,000,000 to Rp15,000,000
- Morfem
- Rp3,000,000 to Rp6,000,000
Monthly estimate. Agents are costed at Rp800 to Rp1,000 per talk minute, based on Bandung and Jakarta call center salaries plus overheads. Morfem is costed at Rp200 to Rp400 per minute, before phone charges. We set the final price after the pilot.
Collection rules are built into every call
We write these rules as code, not as suggestions to the model. The assistant cannot break them.
- Calls only Monday to Saturday, 08:00 to 20:00 local time, as the OJK collection rules require.
- The assistant says it is a voice assistant in its first sentence.
- It mentions the bill only after the customer confirms who they are.
- The customer can ask for a person at any time.
- Data is processed on servers in Indonesia, under the Personal Data Protection Law.
Start with a 30-day pilot
We write the scripts in your customers' language and pick 500 customers at random. Half get the call in Indonesian, half in Sundanese or Javanese. After 30 days we compare on-time payments between the two groups. You decide whether to continue.