thusinh1969/BrightTO-SV-V1.2-LMF_E0-PROD

🏆 BrightoSV Speaker Verification V1.2-LMF

COMMERCIAL SOTA • LARGE MARGIN FINE-TUNED

Bank-Grade Voice Identity Verification (Xác thực Định danh Giọng nói cấp Ngân hàng)

---

BrightoSV Speaker Verification V1.2-LMF is an upgraded release of V1.2, fine-tuned with Large Margin Fine-tuning (LMF) to push bank-grade operating points further. All 8 evaluation configurations surpass the V1.2-E7 baseline — delivering the new production SOTA for offline voiceprint verification.

(BrightoSV V1.2-LMF là bản nâng cấp của V1.2, tinh chỉnh bằng Large Margin Fine-tuning (LMF) để cải thiện hiệu suất tại các điểm vận hành ngân hàng. Toàn bộ 8 cấu hình đều vượt baseline V1.2-E7 — đạt SOTA mới cho xác thực giọng nói offline.)

---

📈 V1.2-LMF vs V1.2 — What Improved

Config	Metric	V1.2 (E7)	V1.2-LMF	Δ
🏦 4s / 5-Enroll	EER	1.184%	1.174%	−0.010% ✅
🏦 4s / 5-Enroll	FRR @ FAR 0.1%	3.18%	3.11%	−0.07% ✅
🏦 4s / 5-Enroll	FRR @ FAR 0.01%	6.65%	6.44%	−0.21% ✅
🏦 4s / 5-Enroll	Tail Gap 5%	+5.3463	+5.3870	+0.041 ✅
🏦 4s / 3-Enroll	FRR @ FAR 0.01%	8.60%	8.51%	−0.09% ✅
📱 2s / 5-Enroll	FRR @ FAR 0.01%	12.58%	12.13%	−0.45% ✅
📱 2s / 3-Enroll	FRR @ FAR 0.01%	15.91%	15.44%	−0.47% ✅

All 8 configurations improved. At production scale, every 0.1% FRR reduction means thousands fewer genuine users rejected per day.

(Toàn bộ 8 cấu hình đều cải thiện. Ở quy mô production, mỗi 0.1% giảm FRR đồng nghĩa hàng ngàn người dùng ít bị từ chối mỗi ngày.)

---

🏆 Key Performance Indicators

🏦 Bank-Grade Benchmarks (4-Second QA Gate, 4s/2s Windows)

Evaluated on 10,000,000 positive + 10,000,000 negative pairs with strict bank-grade QA (Audio ≥ 4s, SNR ≥ 10dB, Speech Ratio ≥ 15%):

Metric (Chỉ số)	3-Enrollment	5-Enrollment	Significance (Ý nghĩa)
EER	1.454%	1.174% 👑	🏆 Commercial SOTA. Sub-1.2% on diverse multilingual test set. (Tỷ lệ lỗi cân bằng dưới 1.2%).
FRR @ FAR=0.1%	4.26%	3.11% 👑	✅ Bank-Grade Achieved. Under 5% at 1-in-1,000 impostor protection. (Đạt chuẩn ngân hàng tại FAR 0.1%).
FRR @ FAR=0.01%	8.51%	6.44% 👑	🔒 Maximum Security. 1-in-10,000 impostor protection for high-value transactions. (Bảo vệ 1/10,000 cho giao dịch giá trị cao).
Threshold @ FAR=0.1%	2.5374	2.6330	🎯 Production threshold at balanced operating point.
Threshold @ FAR=0.01%	4.5337	4.6411	🎯 Production threshold at maximum security.
Tail Gap 5%	+4.4937	+5.3870 👑	🛡️ Full Separation. 95th percentile genuine scores fully above worst impostor region. (Phân tách hoàn toàn vùng đuôi).
Latency ⚡	—	< 60ms	Real-time processing on consumer GPU. (Xử lý thời gian thực).

📱 Extended Coverage Benchmarks (2-Second QA Gate, 2s/1s Windows)

For scenarios requiring shorter audio — mobile apps, call centers, IoT:

Metric (Chỉ số)	3-Enrollment	5-Enrollment	Significance (Ý nghĩa)
EER	2.294%	1.687% 👑	🏆 Best-in-Class. Strong accuracy even with 2-second audio. (Độ chính xác cao ngay cả với audio 2 giây).
FRR @ FAR=0.1%	7.81%	5.90%	✅ Practical Consumer UX. Manageable retry rate for mobile. (Tỷ lệ xác thực lại phù hợp ứng dụng di động).
FRR @ FAR=0.01%	15.44%	12.13%	🔒 High Security on Short Audio. Viable for multi-factor deployments. (Bảo mật cao trên audio ngắn).
Threshold @ FAR=0.1%	2.1329	2.2478	🎯 Production threshold at balanced operating point.
Threshold @ FAR=0.01%	4.1170	4.1908	🎯 Production threshold at maximum security.
Tail Gap 5%	+2.4170	+3.2310	🛡️ Positive Separation. Genuine and impostor tails clearly separated. (Vùng đuôi tách biệt rõ ràng).

---

📊 Evaluation Methodology — Statistical Rigor

Unlike many speaker verification benchmarks that report on small test sets, BrightoSV V1.2-LMF is evaluated at industrial scale:

Aspect	Detail
Positive pairs	10,000,000 (same-speaker, cross-utterance)
Negative pairs	10,000,000 (different-speaker, balanced 1:1)
Total scored pairs	20,000,000
Unique speakers	3,900+ in evaluation set
Multi-enrollment	3 and 5 enrollment utterances, mean-aggregated
Score normalization	AS-NORM with 600,000+ speaker cohort
Quality fusion	QMF (Quality Metric Fusion) — compensates for speaker-specific and duration offsets
QA gates	5-gate bank-grade: Duration, Clipping, RMS Energy, Speech Ratio, SNR

This scale ensures that operating points at FAR=0.01% (1 in 10,000) are backed by actual counts of 1,000 impostor threshold crossings, not statistical extrapolation.

(Quy mô này đảm bảo các chỉ số tại FAR=0.01% được xác thực bởi 1,000 mẫu vượt ngưỡng thực tế, không phải ngoại suy thống kê.)

---

🔬 What is Large Margin Fine-Tuning (LMF)?

LMF is a post-training optimization stage applied after the base model converges. Starting from the V1.2-E7 checkpoint:

Aspect	Base Training (V1.2)	LMF (V1.2-LMF)
Angular margin	0.35	0.40 (+14.3%)
Data selection	Mixed duration (≥2s)	≥4s only (high-information)
Window strategy	3 windows	2 windows (first/last, cleaner)
Learning rate	Full training LR	Finetune LR (1000× lower backbone)
Optimizer	From scratch	Preserved (7 epochs of gradient history)
Scheduler	Continuous	Fresh cosine cycle

The increased margin forces the model to push genuine and impostor embeddings further apart on the hypersphere. Combined with high-quality 4s-only data and preserved optimizer state, this delivers targeted improvements at the most critical operating points.

(LMF tăng angular margin, buộc model đẩy embedding genuine và impostor xa hơn trên hypersphere. Kết hợp với data 4s chất lượng cao và optimizer state được giữ lại, mang đến cải thiện có mục tiêu tại các điểm vận hành quan trọng nhất.)

---

🌍 Multilingual Training — Global Voice Coverage

BrightoSV V1.2-LMF inherits the full multilingual foundation of V1.2:

Language	Coverage	Notes
🇻🇳 Vietnamese	★★★★★	Primary language. Extensive dialect coverage (Northern, Central, Southern). (Ngôn ngữ chính, bao phủ đầy đủ phương ngữ Bắc, Trung, Nam).
🇬🇧 English	★★★★★	Primary language. Multiple accents (US, UK, AU, Indian, Singapore)
🇨🇳 Chinese	★★★★★	Primary language. Mandarin and regional variants
🇰🇷 Korean	★★★★☆	Native speaker corpus
🇩🇪 German	★★★★☆	European language coverage
🇫🇷 French	★★★★☆	Including African French variants
🇳🇱 Dutch	★★★★☆	European language coverage
🇯🇵 Japanese	★★★★☆	Native speaker corpus
🇸🇦 Arabic	★★★★☆	Multiple dialect coverage
🇮🇩 Indonesian	★★★☆☆	Southeast Asian coverage

Key principle: Speaker identity is carried by vocal tract shape, pitch dynamics, and articulatory patterns — these are language-independent. A speaker can enroll in Vietnamese and verify in English.

(Nguyên tắc then chốt: Định danh người nói mang tính phổ quát, không phụ thuộc ngôn ngữ.)

---

🛡️ Robustness — Augmentation & Real-World Resilience

🔊 Noise Resilience

Category	Examples	Goal
🏙️ Urban	Street noise, sirens, traffic, construction	On-the-go verification
🏠 Domestic	TV/radio background, appliances, children	Work-from-home reliability
🗣️ Babble	Crowd noise, overlapping speakers, cafeteria	The hardest scenario — solved
⛈️ Natural	Wind, rain, thunder	Outdoor stability
🐾 Biological	Coughing, sneezing, baby crying	Disentangle speaker from artifacts

📡 Channel & Codec Resilience

Codec / Channel	Simulation
GSM / AMR	Mobile telephony compression
VoIP (Zalo, WhatsApp)	Internet calling artifacts
MP3 / AAC / OGG	Lossy compression at various bitrates
Microphone variance	Laptop, phone, headset, far-field
Room acoustics	Reverb, echo, room impulse response

🎛️ SpecAugment

Time masking applied on WavLM hidden states during training to prevent overfitting, forcing the model to learn robust speaker representations from partial temporal information.

Result: The model maintains accuracy down to 10 dB SNR — equivalent to speaking in a moderately noisy café. Below 10 dB, the QA gate rejects the audio, protecting against unreliable decisions.

---

🎯 Production Deployment

Three Security Levels

Level	Min Audio	Enrollment	FAR Options	Use Case
🏦 bank_strict	≥ 4.0s	5 samples	0.1% / 0.01%	High-value banking, government (Ngân hàng giá trị cao)
🏛️ bank_flex	≥ 4.0s	3 samples	0.1% / 0.01%	Standard banking, telecom (Ngân hàng tiêu chuẩn)
📱 consumer	≥ 2.0s	3 samples	0.1% / 0.01%	Mobile apps, call centers, IoT (Ứng dụng di động)

Scoring Pipeline

Audio → QA Gate (5 checks) → Windowed Embedding Extraction → AS-NORM (600K cohort) → QMF → Decision

Component	Detail
Embedding	512-dimensional voiceprint
Score normalization	AS-NORM with top-300 cohort matching
Quality fusion	Cohort Mean Fusion + Duration compensation
Multi-enrollment	Mean-aggregated across sessions
Window strategy	All valid windows from audio, mean-aggregated (matches eval pipeline)
Storage	~2 KB per enrolled speaker

QA Gate — Mandatory Pre-Check

Check	Threshold	Purpose
Duration	≥ 2s or ≥ 4s	Sufficient speech content
Clipping	< 0.1%	No distorted audio
RMS Energy	−45 to −5 dBFS	Proper recording level
Speech Ratio	≥ 15%	Actual speech, not silence
SNR	≥ 10 dB	Acceptable noise level

Without QA gate, tail performance degrades significantly. QA is mandatory for production. (Không có QA gate, hiệu suất vùng đuôi giảm đáng kể. QA là bắt buộc.)

---

⚙️ Technical Specifications

Specification	Value
Model Version	V1.2-LMF (Large Margin Fine-tuned SOTA)
Base Checkpoint	V1.2-E7
LMF Method	Angular margin 0.35 → 0.40, 4s-only data, finetune LR
Parameters	316M (High-Capacity Self-Supervised Backbone)
Embedding Dimension	512
Input Sample Rate	16kHz (Auto-resampling supported)
Input Formats	WAV, FLAC, MP3, OGG, M4A
Output	512D L2-normalized embedding
Backends	PyTorch, ONNX, HuggingFace

🚀 Hardware & Performance

Specification	Value
GPU Support	NVIDIA T4, A10, A100, H100, L4
CPU Support	Intel Xeon, AMD EPYC (via ONNX)
Inference Latency	< 60ms (GPU) / < 400ms (CPU ONNX)
Model Size	~1.2 GB
Batch Processing	Supported
Deployment	Fully offline after initial download

---

🔒 Privacy & Security

Aspect	Implementation
Audio Retention	Zero. Audio processed in RAM, immediately discarded. (Không lưu audio).
Voiceprint	512 numbers. Non-reversible — cannot reconstruct voice. (Không thể tái tạo giọng nói).
Deployment	On-premise or private cloud. No external calls. (Triển khai nội bộ, không gọi ra ngoài).
Compliance	GDPR, PDPA, PCI-DSS ready
Data Sovereignty	100% local processing. Your data never leaves your infrastructure.

---

🤝 Combined with Anti-Spoofing

For maximum security, deploy BrightoSV Speaker Verification alongside BrightoSV Anti-Spoofing V1.5:

Audio → Anti-Spoof Check (Is this a real voice?) → Speaker Verify (Is this the right person?) → Decision

Layer	Model	Purpose
Layer 1	Anti-Spoof V1.5	Reject deepfakes, replay attacks, TTS
Layer 2	Speaker Verify V1.2-LMF	Confirm speaker identity

This dual-layer architecture provides defense-in-depth: even if a sophisticated deepfake passes liveness detection, it must still match the enrolled voiceprint — and vice versa.

(Kiến trúc hai lớp cung cấp phòng thủ theo chiều sâu: ngay cả khi deepfake vượt qua kiểm tra liveness, vẫn phải khớp voiceprint — và ngược lại.)

---

📈 Roadmap

Version	Status	Highlight
V1.2	✅ Released	Commercial SOTA baseline — EER 1.184%
V1.2-LMF	🟢 Current	New SOTA — EER 1.174%, FRR@0.01% 6.44%
V1.5-SE	🔵 In progress	Full retrain — clean data, new schedule, targeting FRR@0.01% ≤ 5%
V2.0	🟡 Planned	Next-generation architecture

---

📞 Access & Licensing

This model is Private and available exclusively for enterprise partners under NDA. (Model nội bộ, chỉ cung cấp cho đối tác Doanh nghiệp ký NDA.)

Thương mại & Triển khai

• License trọn gói hoặc qua API
• Hỗ trợ tích hợp theo yêu cầu (triển khai, tối ưu hiệu năng, giám sát chất lượng)
• Công ty Cổ phần SphinX (sphinxjsc.com) được giao quyền đóng gói, cung cấp API và phân phối

Bản quyền & License

Thương mại / Proprietary. Việc sử dụng, phân phối lại hoặc tạo bản phái sinh cần có chấp thuận bằng văn bản từ BrighTO Technology.

Liên hệ

Purpose	Contact
Commercial Licensing	nguyen@brighto.ai, nghia@brighto.ai
API & Distribution	duc@sphinxjsc.com (SphinX JSC)
Technical Inquiries	nguyen@hatto.com

---

🏆 BrightoSV Speaker Verification V1.2-LMF

New SOTA • Large Margin Fine-Tuned • Bank-Grade • Offline-Ready

EER 1.174% · FRR 6.44% @ FAR 0.01% · 20M Eval Pairs · 600K Cohort · 9+ Languages

Built in Vietnam 🇻🇳 • Engineered for the World 🌏

---

This model card refers to BrightoSV Speaker Verification V1.2-LMF (Large Margin Fine-tuned Release). This is a direct upgrade from V1.2-E7, with all 8 evaluation configurations improved. All benchmark results are verified on internal evaluation sets comprising 20,000,000 scored pairs across 3,900+ speakers with strict bank-grade QA methodology and AS-NORM score normalization using a 600,000+ speaker cohort.