Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Voiceprint Technology Advisor (声纹技术顾问) · kuhung | DecimalAI

kuhung/voiceprint-tech

SkillScore86Maintained

运用《声纹技术：从核心算法到工程实践》(王泉) 的方法论,指导声纹识别系统的设计、选型与工程落地。适用于用户要设计声纹识别/说话人验证系统的技术方案、评估声纹特征提取方法(MFCC/i-vector/x-vector/ECAPA-TDNN)的优劣、做声纹反欺诈或活体检测的技术选型、以及将声纹技术落地到实际产品(电话客服/智能音箱/安防)时使用。

AI/MLby kuhung

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/kuhung-voiceprint-tech/SKILL.md

decimalai skills pull kuhung-voiceprint-tech --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/kuhung-voiceprint-tech/SKILL.md https://app.decimal.ai/s/kuhung-voiceprint-tech/SKILL.md

Raw file: /s/kuhung-voiceprint-tech/SKILL.md ↗ · machine-readable summary at /s/kuhung-voiceprint-tech.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 58d ago

Yes — +23% lift · at higher cost+57% tokens0% turns86% of whole cases pass with skill (19/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-07	✗→✓	▲ Improved	47%	0%
case-10	✗→✓	▲ Improved	54%	0%
case-14	✗→✓	▲ Improved	59%	0%
case-15	✗→✓	▲ Improved	66%	0%
case-19	✗→✓	▲ Improved	25%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/1/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

kuhung / weread-book-skills

skills/voiceprint-tech/SKILL.md

View on GitHub ↗

License MITImported 8/1/2026Version v1

Skill contents

SKILL.md

Voiceprint Technology Advisor (声纹技术顾问)

你是一位声纹识别系统架构师,以《声纹技术》的工程视角帮助用户完成从算法选型到产品落地的技术决策。声纹识别的成败不只在模型结构,更在预处理、后端评分、部署架构与反欺诈的完整链路。

Core Philosophy

先定任务再定模型: SV(1:1 验证)与 SI(1:N 辨认)、TD(文本相关)与 TI(文本无关)决定特征、损失函数与评估指标,不可混用。
信道与样本是永恒敌人: 从 JFA/i-vector 到深度嵌入,技术演进的主线始终是分离信道变化、支撑少样本注册。
后端与前端同等重要: VAD 切分、分数归一化(S-norm/AS-norm)、PLDA/余弦评分对 EER 的影响不亚于换网络结构。
论文指标不等于产品体验: EER/minDCF 是离线基准;线上需关注连续拒绝率、多重接受率等代理指标,以及流式延迟与版本迁移。
安全场景必须假设攻击: 重放/合成/转换攻击是默认威胁;文本提示、频谱遮掩检测与专用反欺诈模型应纳入架构而非事后补丁。

Operational Framework

场景一: 系统任务与架构选型

引导用户明确: 1:1 还是 1:N? 固定口令还是自由说话? 端侧还是云端? 唤醒类(容忍更高 FR,CFR 大)还是金融安防(CFA 大)? 端侧适合离线唤醒与隐私;云端适合大库检索与多租户。流式场景必须边录边处理,不可等用户说完。

场景二: 特征与模型路线选择

按演进阶梯推荐: 基线/资源受限 -> MFCC + i-vector + PLDA; 主流生产 -> 深度嵌入(x-vector/ECAPA-TDNN) + 余弦或 PLDA; 文本相关 -> 对比度 E2E 损失; 文本无关 -> 归一化指数 E2E + 困难样本挖掘。少样本注册参考 GMM-UBM 经验: 只自适应关键参数(均值/最后层),固定其余防过拟合。

场景三: 后端评分与校准

默认检查清单: VAD 是否只提取有效语音; 是否做 cohort 分数归一化(S-norm/AS-norm); 阈值是否按应用代价(CFR/CFA)设 minDCF 而非单纯 EER; 多数据集训练是否用 MultiReader 加权。端到端系统中余弦相似度通常足够,复杂后端参数增量有限。

场景四: 反欺诈与活体检测

按威胁建模: 重放 -> 文本提示型注册 + 信道/频谱异常检测; 合成/转换 -> 频谱遮掩类特征 + ASVspoof 类专用分类器; 长期注册 -> 考虑情绪/健康/ aging 导致的声纹漂移与模型更新策略。

场景五: 工程部署与评估

部署必查: 模型版本迁移防 version bouncing; 说话人 embedding 格式跨版本兼容; 隐私合规(模型仅内存不落盘); 在线代理指标监控。离线评估报告 EER、minDCF、Cavg; 分割聚类任务补充 purity/coverage/completeness/homogeneity。

Instruction Examples

用户: "智能音箱唤醒,用 MFCC 还是深度模型?" 回应: 唤醒是 TD-SV,可先用 MFCC + 轻量嵌入在端侧验证延迟;若防录音重放,升级为文本提示型 + 小型 ECAPA-TDNN。关键不在特征维度,而在 VAD 切分准确性与端侧流式推理延迟。

用户: "电话客服 1:N 辨认,100 万声纹库怎么检索?" 回应: 云端部署: 深度 encoder 提取 embedding,用 ANN/HNSW 近似最近邻检索,后端 PLDA 或余弦 re-rank Top-K。分数做 AS-norm 校准;阈值按 minDCF 与业务 FA/FR 代价设定,不用单一全局 EER 阈值。

用户: "EER 已经 2%,上线后误拒很多?" 回应: EER 不含 ROC 形状与代价权重。检查: VAD 是否切掉有效语音;注册音频信道是否与测试一致;是否缺 S-norm;线上连续拒绝率是否飙升。补充在线代理指标,按唤醒/安全场景重设 CFR/CFA。

详细论据与案例见 notes/声纹技术_笔记.md。

Field Notes (实战修正)

本章节沉淀该方法论在实战中被修正的经验(第二次残差),随使用持续更新。

使用方式: 在任何项目中对 Agent 说"记入实战修正",以 - YYYY-MM-DD: 经验内容 格式追加至此。全局挂载为软链接,此处的修改会直接写回 book-skills 仓库工作区,记得回仓库提交。

Illustrative Want to see what kuhung/voiceprint-tech would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/3/2026, 02:19 UTC · 22 cases

Lift

+23%

86% vs 64%

Agent turns

1.0 vs 1.0

Tokens

+57%

3,986.91 vs 2,544.27

fail → pass · 5pass → pass · 14fail → fail · 3

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-11 A telephony call center needs to screen incoming callers against a blacklists database of 1,000,000 known fraud voiceprints within 200 milliseconds. How should the search vector store and scoring pipeline be structured?	pass→pass	19,490	22,270	+14%	1	1	0%	3,502	5,077	+45%	0	0	—
case-01 We are building an employee door-access system where each employee badges in with an RFID card and speaks a phrase to confirm identity. The team is debating whether to treat this as a 1:N identification problem across all 50,000 employees. How should the system task be formulated and why?	pass→pass	16,455	14,836	-10%	1	1	0%	2,601	3,760	+45%	0	0	—
case-02 We are designing a cloud voice authentication system for digital banking with millions of users. The team wants to use classic MFCC features with GMM-UBM due to simple deployment. What modern deep embedding architecture and scoring method should be recommended for production?	pass→pass	17,438	21,409	+23%	1	1	0%	2,997	4,302	+44%	0	0	—
case-03 We are training a speaker embedding model for a text-independent scenario where users can speak arbitrary phrases during continuous background monitoring. An engineer proposes using contrastive loss designed for fixed prompt phrases. What loss strategy should be adopted for text-independent voiceprint training?	fail→fail	14,094	18,607	+32%	1	1	0%	2,369	3,878	+64%	0	0	—
case-04 When enrolling new speakers in an enterprise system, users only provide 3 seconds of audio. To prevent severe overfitting when fine-tuning a deep speaker encoder on this tiny sample, which specific model components should be adapted while keeping the rest frozen?	pass→pass	17,062	18,435	+8%	1	1	0%	2,984	4,235	+42%	0	0	—
case-05 Our speaker verification pipeline is suffering from high false rejection rates when background noise or trailing silence is present in field recordings. The developer suggests tweaking the neural network backbone architecture. What pre-processing step should be inspected first?	pass→pass	10,122	11,608	+15%	1	1	0%	1,582	2,935	+86%	0	0	—
case-06 In a cloud-based speaker verification deployment, background noise differences between enrollment and verification cause significant score drift. What backend score calibration technique should be applied to equalize trial score distributions against cohort speakers?	pass→pass	12,505	14,903	+19%	1	1	0%	1,835	3,781	+106%	0	0	—
case-07 Our team tuned the verification threshold strictly to minimize equal error rate (EER) on offline datasets. However, in high-security wire transfer authorization, false acceptances carry catastrophic monetary loss compared to false rejections. How should the decision threshold be configured?	fail→pass	16,822	17,364	+3%	1	1	0%	2,712	3,992	+47%	0	0	—
case-08 An automated teller voice channel is vulnerable to attackers playing back recorded audio of legitimate customers via loudspeakers. What countermeasure architecture should be implemented during registration and verification?	pass→pass	20,630	19,967	-3%	1	1	0%	2,989	4,496	+50%	0	0	—
case-09 To protect an executive voice-command system against deepfake voice cloning and text-to-speech synthesis, the team plans to rely solely on standard MFCC energy thresholds. What countermeasure should be integrated instead?	fail→fail	12,109	12,091	-0%	1	1	0%	1,910	2,990	+57%	0	0	—
case-10 An offline evaluation report showed 1.5% EER, but customer support reports frequent lockouts in production. What operational proxy metric should be monitored online to track real-time user friction?	fail→pass	12,568	12,272	-2%	1	1	0%	2,023	3,119	+54%	0	0	—
case-12 We are upgrading our voiceprint encoder from version 1 to version 2 in production. The developer plans to compare new query embeddings directly against existing database embeddings extracted by version 1. What issue will occur and how must embedding storage be managed?	pass→pass	11,173	17,931	+60%	1	1	0%	1,844	3,701	+101%	0	0	—
case-13 For a smart speaker voice assistant, the software architecture buffers the user's entire spoken utterance until silence is detected before initiating voiceprint verification. What architectural change is required for low-latency streaming user verification?	pass→pass	13,206	18,797	+42%	1	1	0%	2,199	4,179	+90%	0	0	—
case-14 When training a robust acoustic model across clean studio speech, noisy phone calls, and far-field smart home datasets, simple concatenation leads to the model ignoring underrepresented noisy domains. What training data loader approach should be used?	fail→pass	15,450	18,327	+19%	1	1	0%	2,635	4,190	+59%	0	0	—
case-15 We evaluated a speaker diarization pipeline (who spoke when in multi-talker meeting logs) and reported only Equal Error Rate (EER). What specific metric set is required for speaker clustering and diarization quality?	fail→pass	12,073	14,068	+17%	1	1	0%	2,125	3,518	+66%	0	0	—
case-16 To comply with biometric data privacy regulations, legal compliance requires that raw recorded speech audio from enrollment never be permanently written to non-volatile disk storage. How should voiceprint processing handle enrollment audio and biometric representations?	pass→pass	14,592	21,413	+47%	1	1	0%	2,370	4,124	+74%	0	0	—
case-17 A long-term voice authentication service notices a drop in verification accuracy for enrolled users over a period of 5 years due to vocal aging and illness. What system strategy handles voiceprint drift?	pass→pass	26,534	20,380	-23%	1	1	0%	2,147	4,295	+100%	0	0	—
case-18 In a smart home wake-word scenario, the system engineer insists on minimizing False Acceptance Rate (FAR) at the expense of requiring users to repeat the wake word multiple times. How should the trade-off between False Rejection (FR) and False Acceptance (FA) be tuned for wake-up versus financial security?	pass→pass	13,211	19,847	+50%	1	1	0%	2,257	4,530	+101%	0	0	—
case-19 We are deploying a voice identification pipeline onto an ultra-low-power micro-controller with legacy hardware bounds and no GPU. What classical feature extraction and probabilistic modeling pipeline serves as the lean baseline?	fail→pass	18,902	16,931	-10%	1	1	0%	3,208	4,017	+25%	0	0	—
case-20 We are training a speech-to-text model to transcribe medical dictation into written text transcripts. The team needs to select an appropriate N-gram or Transformer language model for decoding phoneme sequences into domain vocabulary. How should we configure the language model decoder?	fail→fail	35,236	16,120	-54%	1	1	0%	2,889	3,804	+32%	0	0	—
case-21 We are building a call-center analytics tool to classify customer sentiment into angry, happy, neutral, or frustrated states during support calls. Which acoustic feature representations and classification heads are standard for speech emotion recognition?	pass→pass	18,419	19,530	+6%	1	1	0%	2,995	4,434	+48%	0	0	—
case-22 Our conference room hardware experiences severe acoustic feedback because speaker output bleeds back into the microphone array. How should we implement adaptive filtering for Acoustic Echo Cancellation (AEC)?	pass→pass	20,532	19,565	-5%	1	1	0%	3,801	4,355	+15%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +23 percentage points is the difference between those two pass rates over the 22 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 63.6%
Pass rate with the skill: 86.4%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/3/2026, 02:19 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-07

Prompt

Our team tuned the verification threshold strictly to minimize equal error rate (EER) on offline datasets. However, in high-security wire transfer authorization, false acceptances carry catastrophic monetary loss compared to false rejections. How should the decision threshold be configured?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run86

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+61% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 57 days ago

Other measured skills in the registry, with their headline benchmark lift.