vdavidr/textToAudioTTS-voxcpm2-audiobook-champion

VoxCPM2 Audiobook Champion Artifacts

This repository contains the current textToAudioTTS audiobook champion artifacts from the 2026-04-24 evaluation run.

Important: this is not a re-upload of the openbmb/VoxCPM2 base weights. The generated samples and quality-gate scripts here document the working inference recipe and measured output quality. Use openbmb/VoxCPM2 as the base model.

Current Winner

• Model path tested: openbmb/VoxCPM2
• Mode: plain TTS, no voice clone
• Test input: long-form English audiobook passage
• Duration: 224.48s
• Sample rate: 48000 Hz
• Coverage: 1.0000
• WER: 0.0137
• Prefix WER: 0.0137
• Hypothesis/reference words: 949 / 949

Included Files

• samples/CHAMPION_voxcpm2_original_long_audiobook_test.wav: full winning sample.
• samples/CHAMPION_voxcpm2_original_long_audiobook_test_loudnorm.wav: loudness-normalized full sample.
• samples/CHAMPION_voxcpm2_original_long_audiobook_test_loudnorm_first60s.wav: quick listening cut.
• quality_gate/asr_quality_audit_longform_top_candidates_2026-04-24.json: Whisper ASR audit for top candidates.
• quality_gate/longform_top_candidates_quality_gate_2026-04-24.json: deterministic pass/fail gate and selected winner.
• scripts/audit_tts_asr_quality.py: local ASR audit utility.
• scripts/select_tts_champion.py: quality gate and champion selector.

Quality Gate

Default acceptance criteria:

• Minimum duration: 15s
• Minimum coverage: 0.90
• Maximum WER: 0.05
• Maximum prefix WER: 0.08

The current winner passed. The Inworld TTS training experiment from the same run was rejected because the final generated samples were short and ASR-incoherent.