YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Tajik TTS Listening Pack — voice selection for the Tajik VoiceChat (full-duplex) project
Saidzoda Lab, 2026-08-07. Round-trip WER = synthesized audio transcribed back by our Whisper-tg (9.81 WER) judge vs the source text; real-human-speech floor is ~10.8%. Numbers from files.
| Dir | Model/voice | Round-trip WER | Role considered |
|---|---|---|---|
| 1_orpheus_spk8_CHAMPION | Orpheus-3B single-voice spk8 | 15.91% | AGENT voice (the assistant) |
| 2_orpheus_spk4 | Orpheus-3B single-voice spk4 | 16.67% | user voice |
| 3_orpheus_multivoice_m1 | Orpheus-3B 19 narrators (warm-start) | 19.13% | user-voice diversity pool |
| 4_chatterbox | Chatterbox multi-spk | 43% | closed negative (audible reference) |
| 5_voxcpm_v2 | VoxCPM2 LoRA 23k steps | 93.11% | closed negative (audible reference) |
Each dir: wav samples + transcripts.jsonl (audio_filepath, text) so pronunciation can be judged against the source text. Please rate 1/2/3 for: naturalness, pronunciation accuracy, and suitability as (a) agent voice, (b) user voices, for the synthetic duplex dialogue dataset.
- Downloads last month
- 6