Kola Voices
Voices and speech recognition for Nigerian languages.
Kola Voices speaks and understands Nigerian English, Pidgin, Yoruba, Igbo and Hausa. We test it on phone-quality audio.
Ẹ n lẹ́, Kola ni mo jẹ́, mo ń pè yín láti báǹkì. Owó tí ẹ fi ránṣẹ́ ti lọ.
The problem
Popular speech models do badly on Nigerian languages over the phone.
Yoruba, Hausa and Igbo fall apart on a phone line.
Words go missing, and some languages are not supported at all.
40 held-out clips per language, passed through a simulated phone line. Lower is better, and above 1.0 means the model added words. Pidgin counts spelling differences as errors.
The solution
One model that speaks. One that listens.
Listen.
Hear a sentence in five languages.
Pick a language and a voice, then press play. The clips on this page are recordings of the voices, so they play at once.
@Nigerian English Hello, this is Kola, calling from the bank. Your transfer was successful.
Voices.
Five languages, a female and a male voice for each. Pick a language and press play.
Ẹ n lẹ́, Kola ni mo jẹ́, mo ń pè yín láti báǹkì. Owó tí ẹ fi ránṣẹ́ ti lọ.
- Nigerian English
- Pidgin
- Yoruba
- Igbo
- Hausa
- FemaleMale
These are recordings, so they play at once. Typing your own sentence is not open yet.
Ears.
Speech recognition tested on phone calls.
We trained it on about 70 hours of Nigerian speech. Most of that audio was first passed through a phone-line filter. It writes down Nigerian English, Pidgin, Yoruba, Igbo and Hausa, and there is no per-minute fee.
Results.
How it compares with ElevenLabs Scribe.
Our model against ElevenLabs Scribe v2 on the same phone-quality clips. Words wrong per 100, so lower is better.
Words wrong per 100 on 40 held-out clips per language, passed through a simulated phone line (8 kHz). Hausa is 6 points behind, and we are adding more Hausa data. Pidgin counts spelling differences as errors.
Nigeria has hundreds of languages. We have five so far.
How it works
How we train it.
Open data and recordings
We start with openly licensed Nigerian speech and add recordings from native speakers who agree to it. Each recording goes through six checks before we keep it.
- LengthNot too short, not too long
- LoudnessNot too quiet, not too loud
- DistortionNo clipping
- NoiseA quiet enough room
- Cut-off wordsStarts and ends cleanly
- Wrong wordsOur speech recognition reads it back
A phone-line filter
About 70% of the audio the recognition model learned from was first passed through a filter that copies a phone line: 8 kHz, narrow band, phone compression.
Streaming speech
The voices start speaking while the sentence is still being generated, so the first sound arrives in under a second.
Who we are
A small team in Lagos.
Kola Voices is made by KOLAPPTECH LTD. We record, train and test the models ourselves. To talk to us, write to team@kolapp.tech.
Where things stand
What works now, and what is next.
Working now
- Ten voices in five languages
- Speech recognition within a point of Scribe in four languages
- A recording page for native speakers
Next
- Native-speaker recordings: English and Pidgin first, then Igbo and Hausa
- Typing your own sentence, and speaking to the recognition model
- More Hausa data for the recognition model
- Accent training for the English and Pidgin voices
- A live phone-call path, tested side by side before it takes real calls
Questions
Frequently asked questions
Which languages does it speak and understand?
Nigerian English, Nigerian Pidgin, Yoruba, Igbo and Hausa. Each language has a female and a male voice, and the speech recognition handles all five.
Does it work on a phone line?
Yes. About 70% of the audio the recognition model learned from was first passed through a filter that copies a phone line (8 kHz, narrow band, phone compression). Every score on this page was measured on audio like that, not on clean studio recordings.
How does it compare with ElevenLabs Scribe?
On our test clips we are within a point of Scribe in Nigerian English, Igbo, Pidgin and Yoruba, and 6 points behind in Hausa, where we are adding data. The table above has the numbers and how they were measured.
Whose voices are these?
They are trained on openly licensed Nigerian speech, and we add recordings from native speakers who agree to it. Our recording page checks every take for noise, cut-off words and wrong words.
Can I type my own sentence?
Not yet. The clips on this page are recordings, so they play at once and need no graphics card. Your own sentence and your own audio need a graphics card, and we are finishing that part first. To hear when it opens, write to team@kolapp.tech.
Is it open source? Can I use it?
Not yet. The trained models stay private while we finish checking the licence of every data source. To talk about using them, write to team@kolapp.tech.
Hear it.
Pick a language and press play. Typing your own sentence and speaking to the recognition model are not open yet. Write to us if you want to know when they are.