Metric-AI introduced OhanTTS and the OhanASR family for Armenian on November 20. The synthesizer returns audio incrementally but receives the entire text upfront. Recognition has separate models for recordings and incoming audio. The releases use CC BY-NC 4.0: downloadable weights do not imply unrestricted commercial permission.

OhanTTS also supports other languages and a reference voice. The team discloses pronunciation, long-generation and language-switching limitations. OhanASR’s published results differ between recorded and live speech; combining those results into one accuracy promise would be misleading.

Why separate components matter

For a language with fewer ready-made tools, choice at each stage is useful: turning sound into text and turning an answer back into sound. It lets a developer identify where a mistake originates. A name misunderstood during recognition and a name mispronounced during synthesis require different corrections.

The two meanings of streaming also need separation. Audio starting quickly after receipt of a completed text does not establish the wait while another model is still composing that text. In a conversational system, total delay includes the whole path rather than only its fastest component.

A proposed pilot for a language project

We would begin with short recordings from consenting participants: a quiet room, street noise, names and numbers. Each recording should have a checked transcript. Separately assess whether the synthesized answer preserves those same names and values. This is a proposed future pilot, not an experiment conducted by SIExpo.

For a service offering human assistance, correction matters. After misrecognizing a number, the application should let the user repeat it and show what was actually recorded. Collected corrections help evaluation only when access and retention conditions are clear.

For reference voices, we would use recordings only with the speaker’s permission. Consent for experimentation should also specify whether the generated voice may be published. The model’s license and a person’s consent answer separate questions; satisfying one does not establish the other.

We would keep a small error log organized by category rather than immediately replacing every component after a bad result. A transcription error, a reasoning mistake and a playback fault can sound similar to a listener but belong to different parts of the system. That distinction makes subsequent changes easier to evaluate.

A commercial team’s first step is clarifying the applicable license. Non-commercial researchers receive concrete artifacts they can inspect. In either case, quality should be evaluated against the actual language material and operating conditions of the intended service.