nsanku TTS Benchmark

How the score is calculated. Each synthesised clip is graded two ways and the two are averaged:

• Accuracy = 1 − CER (intelligibility). An ASR judge — the best model for that language in the nsanku ASR benchmark — transcribes the clip, and the transcript is compared with the sentence that was synthesised. CER (shown as a percentage, lower is better) is the share of characters wrong; Accuracy is its complement.
• SBS (SpeechBERTScore, acoustic similarity). The clip is compared with a real human recording of the same sentence (microsoft/wavlm-large, layer 6). Higher is better.
• Composite = (Accuracy + SBS) / 2, higher is better. Example: CER 6.49% and SBS 0.7126 give Accuracy 0.9351 and Composite (0.9351 + 0.7126) / 2 = 0.8238.

Failed clips are not dropped. If a model produces no usable output for a sentence (empty audio, audio past the judge's 39 s limit, or no synthesis), that sentence counts as CER 100% and SBS 0 and is shown in the Failed column, so a model cannot score better by failing on hard sentences. Each model gets one attempt per sentence.
Every sample comes from ghana-speech-eval; a language with several text sources (domains) is sampled evenly across them. The domain columns show the Composite computed on that source alone.