Speech AI Development
Speech systems tested on how people here actually speak.
What is speech AI?
Speech AI converts spoken audio into text or responds to spoken instructions, with accuracy that depends heavily on accent, language mixing, background noise and call quality.
Speech recognition is presented as solved, and it largely is for clear American or British English in a quiet room. That is not the condition most Kenyan businesses record in.
Three things reduce accuracy sharply here. Kenyan English differs from the accents most models were trained on, people switch between English and Swahili mid sentence, and call audio is compressed and noisy.
None of that makes speech systems impractical. It means the honest number comes from testing on your own recordings rather than from a vendor benchmark.
Why we ask for your worst recordings
Duara builds speech systems from Nairobi for call centres, field operations and businesses whose customers prefer speaking to typing.
We test on your recordings before quoting, including the noisy ones. Published accuracy figures are measured on audio that sounds nothing like a Kenyan call centre, and quoting them at you would set an expectation the deployment cannot meet.
Who this is for
- Calls are transcribed or summarised by hand
- Your customers prefer speaking to typing
- Field staff cannot type while working
- Existing speech tools perform badly on your audio
Industries we serve
- Call centres and BPO
- Financial services
- Healthcare providers
- Field services and logistics
- Insurance claims
- Agriculture extension services
Problems this solves
These are the conditions that separate a speech demonstration from a working deployment.
- Accents outside the training data. Models trained mostly on American and British speech are measurably less accurate on Kenyan English.
- Language switching mid sentence. People mix English and Swahili naturally, and most systems handle one language at a time.
- Call audio quality. Telephone audio is compressed and noisy, which is a materially harder problem than a headset in an office.
- Manual transcription and summarising. Calls are written up by hand afterwards, which is slow and captures less than people assume.
What we deliver
Speech systems tested on your own audio, tuned for local accent and language mixing, with uncertain passages flagged rather than guessed, and a defined route to a person.
Speech to text tuned to your audio conditions
Handling of English and Swahili including mixing
Accuracy measured on your own recordings
Uncertain passages flagged rather than guessed
Call summarisation where transcription is not the goal
A route to a person in any voice interface
Features
- Tested on your recordings. Published accuracy figures come from audio nothing like a Kenyan call, so we measure on yours before quoting.
- Language mixing handled. English and Swahili in the same sentence is normal here and the system is built for it rather than tripped by it.
- Uncertain passages marked. Low confidence sections are flagged rather than filled with a plausible guess, which matters most in numbers and names.
- Summaries where transcripts are not the point. Most clients want what the call was about rather than every word, which is a cheaper and more useful output.
- A person always reachable. Any voice interface offers a route to a human, since a caller stuck in a loop is worse than no system.
Benefits
- Calls stop being written up by hand. Transcription and summarising happen automatically and capture more than manual notes do.
- Speech works for your actual users. Testing on local audio means accuracy holds where an imported assumption would not.
- Field staff can record hands free. Speaking works where typing is impractical, which is most of field work.
- Call insight becomes possible. Once calls are text, patterns across them can be seen for the first time.
Technologies we use
- Speech to text. Converting recorded or live audio into text, tuned for your accent and audio conditions.
- Language and code switch handling. Recognising English and Swahili including switching within a sentence.
- Confidence marking. Flagging low certainty passages rather than filling them with a guess.
- Summarisation. Producing what a call was about rather than a full transcript where that is more useful.
- Voice interface design. Spoken interaction with fallback to a person built in from the start.
Accuracy on Kenyan audio is the whole question, and it is answerable only with your recordings. Any figure quoted before hearing your audio is somebody else measurement.
How the process works, step by step
A speech project usually takes five to nine weeks, and it starts with your own audio rather than a specification.
-
Step 1
Free audio assessment
We measure accuracy on your recordings, including the poor quality ones, since those set the real figure.
-
Step 2
Vocabulary supplied
Your product names, place names and terms are added, which improves accuracy more than most tuning.
-
Step 3
Output chosen
We establish whether you need transcripts or summaries, since summaries are cheaper and usually more useful.
-
Step 4
Confidence thresholds set
What gets flagged for a person is agreed, with numbers and names treated more strictly.
-
Step 5
Live with monitoring
Accuracy is tracked after launch, since audio conditions and call routing change over time.
Quality assurance and how we work
The commitments attached to a technology whose accuracy varies by accent and audio.
- We quote accuracy from your audio only. Published figures are measured on recordings nothing like a Kenyan call centre, and repeating them would set an expectation the deployment cannot meet.
- We say when audio quality is the limit. Where recordings are too poor for reliable transcription, improving capture comes before any modelling.
- Uncertain passages are marked, not guessed. A flagged gap is honest. A confident wrong number in a transcript is worse than a blank.
- Voice interfaces always reach a person. A caller trapped in a system with no way out is worse than not having the system at all.
- We test code switching explicitly. Mixing English and Swahili is normal here, and a system only tested on single language audio will disappoint in production.
Frequently asked questions about speech AI
How accurate is it on Kenyan English?
Less accurate than on the accents models were mostly trained on, which is why we measure on your recordings before quoting anything.
Do you quote published accuracy figures?
No. Those are measured on audio nothing like a Kenyan call centre, and repeating them would set an expectation the deployment cannot meet.
Can it transcribe Swahili speech?
Yes, including switching between English and Swahili mid sentence, which we test explicitly rather than assuming.
Why does code switching matter?
Because it is how people here actually speak. A system tested only on single language audio will disappoint in production.
What about poor call quality?
Telephone audio is compressed and noisy, which is materially harder. Where recordings are too poor we say so rather than proceeding.
Can it improve if we supply our terms?
Yes, and supplying product, place and person names usually improves accuracy more than any other tuning.
Do we need full transcripts?
Often not. Most clients want what the call was about, and summaries are cheaper and more useful than every word.
What happens with unclear audio?
The passage is flagged rather than guessed. A confident wrong number in a transcript is worse than a blank.
Can it identify who was speaking?
Yes, where audio quality supports it, which matters for call reviews and dispute handling.
How long does a speech project take?
Five to nine weeks, starting with an assessment on your own recordings.
Can it work in real time?
Yes, though live transcription is harder than recorded and we would test both before committing to it.
Can callers speak to a voice system?
Yes, with a route to a person built in, because a caller trapped in a loop is worse than no system at all.
Does it work for field staff?
Yes, and speaking works where typing is impractical, which covers most field conditions.
Can it analyse calls across a period?
Yes, once calls are text, patterns become visible for the first time, which is often the larger value.
Does audio leave our systems?
That depends on the approach and it is your decision. On premise options exist where recordings cannot leave.
Do we need customer consent to record?
Yes, and that is a legal requirement rather than a technical one. We would raise it early rather than after deployment.
Can it detect sentiment?
Cautiously. Tone detection is less reliable than transcription and we would not build a decision on it.
What if our recordings are too poor to use?
We say so at the assessment. Some audio conditions do not support automation and it is cheaper to know before building.
Who can hear our recordings?
Only those you authorise. Audio stays under your control and is not used to train public models.
What does a speech project cost?
Audio volume and whether you need transcripts or summaries determine it, quoted after the assessment.
Glossary of terms
Speech terms in plain language.
- Speech to text
- Converting spoken audio into written text, also called transcription.
- Code switching
- Moving between languages within a conversation or sentence, which is common in Kenya.
- Word error rate
- The standard accuracy measure for transcription, counting insertions, deletions and substitutions.
- Custom vocabulary
- Your own product, place and person names supplied to improve recognition.
- Diarisation
- Identifying who spoke which part of a recording.
- Confidence marking
- Flagging passages the system is unsure about rather than guessing.
Related services in AI Development
Talk to us about speech AI
Send us twenty recordings including your worst line quality. That assessment answers the accuracy question properly.
Start your project