Ask a global speech service to transcribe an Azerbaijani call and you usually get back fluent, confident text — in Turkish. Chinar is built for Azerbaijani from the ground up, in three production models, all writing punctuation and capitals and labelling speakers: Chinar-F for every call at volume, Chinar-L for documents, Chinar-P for sites without a GPU.
Ask a global speech service for an Azerbaijani transcript and it returns fluent Turkish. It looks like a working transcript — and unless someone on your team reads Azerbaijani, nobody notices until the text reaches a report.
Two major 2026 speech services tested head to head on the same audio — one commercial, one open — produced Azerbaijani output that was effectively unusable. They do not support the language at all.
Many engines learn from clean, read-aloud recordings and fall apart on a real conversation: background noise, interruptions, two people talking at once, ordinary phone-line quality.
Transcribing every call through a cloud service means paying by the hour and handing customer conversations to a third party that keeps its own copy.
Not a multilingual model with Azerbaijani bolted on, and not Turkish with an accent. The language is the design — which is the difference between a transcript you can act on and one that merely looks finished.
A compact model trained on real Azerbaijani call-centre recordings at phone quality, and our most accurate model on real calls: 23% fewer word errors than a commercial cloud speech service on the same call recordings. It writes punctuation and capitals and labels who is speaking. On the public FLEURS benchmark it gets 88.6 words in 100 right with just 24 million parameters. An hour of audio takes about 1 min 40 s on four CPU cores, so every call gets transcribed rather than a sample.
Writes punctuation, capitals and numbers as digits and labels who is speaking, ready to put in front of a supervisor, an auditor or a customer. 93–94% word accuracy on clear Azerbaijani speech, measured against human-checked transcripts. Runs on GPU.
A compact model that needs no GPU: 26× faster than real time on four CPU cores, with punctuation, capitals and speaker labels. On real calls it is on a par with a commercial cloud speech service — on a server you already own.
Trained on genuine call-centre recordings — background noise, interruptions, people talking over each other, ordinary phone-line quality, and mixed Azerbaijani–Russian speech.
Before you talk to us, test hosted ASR models on your own Azerbaijani audio at chinar.allmaz.az: upload a recording, read the transcript, edit and export it. Free, with no card and no sales call.
Chinar runs on your own infrastructure. Customer conversations never leave your building: no per-hour metering, no third-party retention, no questions about where the audio went.
| Chinar-F | Chinar-L | Chinar-P | |
|---|---|---|---|
| Built for | Every call at volume: QA, search, analytics | Documents: minutes, reports, subtitles | Sites without a GPU: branches, closed networks |
| Accuracy on phone calls | Highest — 23% fewer errors than a commercial cloud service | High, close to Chinar-F | On a par with a commercial cloud service |
| Punctuation, capitals, speaker labels | Yes | Yes, plus numbers as digits | Yes |
| Hardware | CPU or GPU | GPU | CPU or GPU — no GPU needed |
| Cost at volume | Lowest | Medium | Low |
Transcribe every conversation instead of a sample, and read the result in Azerbaijani without translating it first.
Punctuated transcripts fit to sit in a compliance record, produced without sending the recording to a third party.
Run Chinar-F across an archive to power search, analytics and quality monitoring at a cost per hour that makes full coverage affordable.
Interviews and meetings transcribed in Azerbaijani, on infrastructure you control.
In head-to-head testing on the same audio, two major 2026 speech services — one commercial, one open — produced Azerbaijani output that was effectively unusable. Chinar returned usable Azerbaijani transcripts of the same recordings. That gap is not a tuning difference: those services do not support the language at all.
Chinar is trained on genuine call-centre recordings, with the noise, the interruptions and the phone-line quality that come with them. Many engines are trained on read speech and fall apart on a real conversation.
Measured against the cloud services we benchmarked, so a day of recordings is processed in minutes rather than left in a queue.
Chinar-F, Chinar-L and Chinar-P are all trained on Azerbaijani phone audio, so you pick the size for the job rather than the language support: F for volume, L for documents, P for sites without a GPU. All three are in production. None of them bills per minute — the cost does not grow with call volume.
The çinar is the oriental plane tree — the broad, long-lived tree that shades courtyards and village squares across Azerbaijan, old enough that people meet under the same one their grandparents did. It is a local thing, deeply rooted, and it outlasts whatever is planted beside it. The name fits a speech model built for this language rather than borrowed from the one next door.
Chinar runs on your own infrastructure, so customer conversations never leave your building: no audio goes to a third party, nothing is retained outside your network, and no per-hour meter runs against a cloud account. The 93–94% figure is word accuracy for Chinar-L, our GPU model, on clear Azerbaijani speech, measured against human-checked transcripts — noisier audio scores lower, and we would rather state the condition than quote a number without it.
Allmaz is the AI product studio of Smart Solutions, which built and operates Azerbaijan's unified public procurement portal — established by presidential decree and delivered as one of the country's first public-private partnerships in digital government.
Request a demo and we will run Chinar on your own Azerbaijani recordings — Chinar-F for the volume, Chinar-L for documents, Chinar-P where there is no GPU.