File size: 2,263 Bytes
08c51bf e2940c9 08c51bf 49dba1f 11be0d1 08c51bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 | {
"model": "Fasih-TTS-V1",
"base_model": "coqui/XTTS-v2",
"language": "ar",
"variant": "MSA / Fusha, single professional male voice",
"evaluation": {
"metric": "CER (%)",
"asr_judge": "openai/whisper-large-v3 (via faster-whisper)",
"normalization": "diacritics-stripped + orthography-normalized (both hyp and ref)",
"sets": {
"varied_msa": {"clips": 8, "mean_cer": 1.3, "worst_cer": 2.2},
"same_sentence_x4": {"clips": 4, "mean_cer": 2.0, "worst_cer": 2.0},
"long_text_chunked": {"clips": 2, "mean_cer": 0.8, "worst_cer": 0.9},
"hard_stress": {"clips": 6, "mean_cer": 2.1, "worst_cer": 8.2},
"human_originals_floor": {"clips": 8, "mean_cer": 1.8, "worst_cer": 4.8}
},
"summary": {
"tts_mean_cer": 1.3,
"human_floor_mean_cer": 1.8,
"gap_points": -0.5,
"instability_events": 0
}
},
"efficiency": {
"gpu": "1x NVIDIA RTX 2080 Ti (Turing, FP32)",
"rtf": 0.60,
"streaming_time_to_first_audio_ms": 675,
"output_sample_rate_hz": 24000
},
"training": {
"data_clips": 1297,
"data_hours": 2.4,
"precision": "fp32",
"best_val_loss": 2.622,
"diacritization": "CATT encoder-decoder"
},
"silma_benchmark": {
"source": "silma-ai/opensource-arabic-tts-benchmark",
"split": "msa",
"sentences": 10,
"judges": ["openai/whisper-large-v3", "nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0", "UTMOS22 (naturalness)"],
"note": "WER=intelligibility (lower better); UTMOS=naturalness proxy (higher better); SILMA's own benchmark is human-auditory.",
"results": [
{"model": "Fasih-TTS-V1 (ours)", "wer_whisper": 6.5, "wer_nemo": 2.5, "utmos": 3.16},
{"model": "xtts (base)", "wer_whisper": 10.3, "wer_nemo": 2.5, "utmos": 2.99},
{"model": "chatterbox", "wer_whisper": 12.8, "wer_nemo": 5.4, "utmos": 3.20},
{"model": "silma_tts", "wer_whisper": 11.1, "wer_nemo": 5.8, "utmos": 3.15},
{"model": "omnivoice", "wer_whisper": 15.3, "wer_nemo": 7.3, "utmos": 3.62},
{"model": "habibi_specialized", "wer_whisper": 21.9, "wer_nemo": 23.3, "utmos": 2.33}
],
"summary": "Fasih is top-tier intelligibility (best-or-tied WER across both ASR judges); mid-pack naturalness (UTMOS #3)."
}
}
|