Split CLI and fix datasets

[BrandonWeng]

Split up the massivceCLI file and fix the benchmarking failure from previous PR. This also makes it so far when its below a threshold it fails.

[github-actions]

VAD Benchmark Results

Performance Comparison

Metric	FluidAudio VAD	Industry Standard	Status
Accuracy	98.0%	85-90%	✅
Precision	96.2%	85-95%	✅
Recall	100.0%	80-90%	✅
F1-Score	98.0%	85.9% (Sohn's VAD)	✅
Processing Time	634.3s (100 files)	~1ms per 30ms chunk	✅

Industry Leaders:

• Silero VAD: ~90-95% F1 (DNN-based, 1.8MB model)
• WebRTC VAD: ~75-80% F1 (GMM-based, fast but lower accuracy)
• Sohn's VAD: 77.5% F1 (traditional approach)
• Modern DNNs: 85-97% F1 (varies by SNR conditions)

[github-actions]

🎯 Single File Benchmark Results

Test File: ES2004a (1049.4s audio)
Overall Result: ✅ PASSED

📊 Accuracy Metrics

Metric	Value
DER (Diarization Error Rate)	18.7%
JER (Jaccard Error Rate)	22.6%
RTF (Real-Time Factor)	0.07x
Speakers Detected	4/4

⏱️ Performance Timing

Stage	Time (s)	% of Total
Model Download	2.334	3.1%
Model Compilation	0.935	1.2%
Audio Loading	0.136	0.2%
Segmentation	15.755	21.0%
Embedding Extraction	56.007	74.5%
Speaker Clustering	0.036	0.0%
Total Processing	75.204	100%

Inference Time: 71.799s (95.5% of total)
Setup Overhead: 3.269s (4.3% of total)

Research Comparison:

• Powerset BCE (2023): 18.5% DER
• EEND (2019): 25.3% DER
• x-vector clustering: 28.7% DER

[Alex-Wengg]

is this just an arbitrary number , seem like the 1 is to represent the min iterations for consistency.

[BrandonWeng]

https://github.com/FluidInference/FluidAudio/pull/24/files/f56b59d56c7e88ad45a95422340985e88fb9a279#diff-eafa60fcbd28ed77f777519cbbb96b9fe180ee35f906349e307cf45d22ec6ba4R61

Not arbitrary. One here just means that we're running it more than once so print out the iteration number

[github-actions]

🎯 Single File Benchmark Results

Test File: ES2004a (1049.4s audio)
Overall Result: ✅ PASSED

📊 Accuracy Metrics

Metric	Value
DER (Diarization Error Rate)	18.7%
JER (Jaccard Error Rate)	22.6%
RTF (Real-Time Factor)	0.04x
Speakers Detected	4/4

⏱️ Performance Timing

Stage	Time (s)	% of Total
Model Download	2.698	5.3%
Model Compilation	0.766	1.5%
Audio Loading	0.099	0.2%
Segmentation	11.471	22.7%
Embedding Extraction	35.578	70.3%
Speaker Clustering	0.025	0.0%
Total Processing	50.637	100%

Inference Time: 47.074s (93.0% of total)
Setup Overhead: 3.464s (6.8% of total)

Research Comparison:

• Powerset BCE (2023): 18.5% DER
• EEND (2019): 25.3% DER
• x-vector clustering: 28.7% DER

[github-actions]

Single File Benchmark Results

Test File: ES2004a (1049.4s audio)
Overall Result: ✅

Accuracy Metrics

Metric	Value
DER (Diarization Error Rate)	18.7%
JER (Jaccard Error Rate)	22.6%
RTF (Real-Time Factor)	0.06x
Speakers Detected	4/4

⏱️ Performance Timing

Stage	Time (s)	% of Total
Model Download	3.179	5.0%
Model Compilation	0.934	1.5%
Audio Loading	0.091	0.1%
Segmentation	13.240	21.0%
Embedding Extraction	45.536	72.3%
Speaker Clustering	0.032	0.1%
Total Processing	63.011	100%

Inference Time: 58.807s (93.3% of total)
Setup Overhead: 4.113s (6.5% of total)

Research Comparison:

• Powerset BCE (2023): 18.5% DER
• EEND (2019): 25.3% DER
• x-vector clustering: 28.7% DER

[github-actions]

VAD Benchmark Results

Performance Comparison

Metric	FluidAudio VAD	Industry Standard	Status
Accuracy	98.0%	85-90%	✅
Precision	96.2%	85-95%	✅
Recall	100.0%	80-90%	✅
F1-Score	98.0%	85.9% (Sohn's VAD)	✅
Processing Time	636.2s (100 files)	~1ms per 30ms chunk	✅

Industry Leaders:

• Silero VAD: ~90-95% F1 (DNN-based, 1.8MB model)
• WebRTC VAD: ~75-80% F1 (GMM-based, fast but lower accuracy)
• Sohn's VAD: 77.5% F1 (traditional approach)
• Modern DNNs: 85-97% F1 (varies by SNR conditions)