Rcarvalo/benchmarkTTS

TTS Model Benchmarks

Benchmark results for various TTS models on French and English, with audio samples sorted worst-to-best by WER.

Summary Results (French - 500 SIWIS phrases)

Model	Samples	WER Mean	WER Median	RTF	Real-time?
Qwen3-TTS 1.7B	500	23.4%	14.3%	1.300	No (0/500)
VibeVoice 0.5B (FT)	500	35.0%	22.9%	0.416	Yes (500/500)
CeSAMe CSM-1B 4-bit	150	69.6%	66.7%	3.246	No

Summary Results (English)

Model	Samples	WER Mean	WER Median	RTF
CeSAMe CSM-1B 4-bit	150	8.7%	0.0%	2.857
Qwen3-TTS 1.7B	300	12.9%	0.0%	1.690

Audio Samples

Each model folder contains a worst_to_best/ directory with all generated audio files ranked by WER (worst first). File format: NNN_werXXX_sampleid.wav

• vibevoice_french/worst_to_best/ - 500 audio samples
• qwen3_tts_french/worst_to_best/ - 500 audio samples

Structure

vibevoice_french/
  results.csv            # Full benchmark results
  worst_to_best/         # Audio ranked by WER (001 = worst)
qwen3_tts_french/
  results.csv
  worst_to_best/
cesame_unsloth_baseline/
  results.csv            # EN + FR baseline (no audio)
qwen3_tts_english/
  results.csv            # EN baseline (no audio)

Evaluation Methodology

• WER: Word Error Rate via OpenAI Whisper API transcription
• RTF: Generation time / audio duration (< 1.0 = real-time capable)
• Benchmark: 500 SIWIS French phrases (seed=42, 15 < len < 300, deduplicated)

Models

• VibeVoice-Realtime-0.5B (Microsoft) - Fine-tuned on SIWIS French → Rcarvalo/vibevoice
• CeSAMe CSM-1B (Sesame) - Unsloth 4-bit quantization
• Qwen3-TTS-12Hz-1.7B (Alibaba) - Base model