Context

The basic TTS mechanics via OpenRouter are already implemented, but a technical issue has been discovered: when sending a single letter for voice generation, OpenRouter returns an HTTP 400 error.

This is critical for the primer, as high-quality audio generation of individual letters, sounds, and short phonetic units is required.

Objective

Research and implement an alternative local TTS pipeline via a local server that can stably generate audio for single letters and other minimal units of the Vietnamese language without OpenRouter limitations.

What needs to be tested

• which local TTS models support the Vietnamese language;
• whether they can correctly synthesize single letters, syllables, and very short strings;
• the quality of tones and phonetics;
• the feasibility of running the model on a local server;
• the API for calls from vietnamguru.ru;
• audio formats and generation latency;
• hardware requirements;
• licenses and usage restrictions;
• the possibility of batch generation of the voiceover dictionary;
• compatibility with the existing audio file and metadata storage system.

Proposed Architecture

vietnamguru.ru → local TTS API → local model → audio file → existing voiceover dictionary/storage.

The local mechanism should be treated as a source audio generator; generated recordings, just like in the main TTS pipeline, should be saved and reused rather than synthesized on every playback.

Acceptance Criteria

Find and verify a local generation method that stably accepts single letters and short phonetic elements of the Vietnamese language and produces sound of sufficient quality for further verification by a native speaker.