Speaker Verification Implementation with Anti-Spoofing

A fintech company whose voice bot processed transfers faced fraud: fake voice commands generated via WaveNet passed verification unchallenged. Losses hit 15% of transaction volume. They needed a speaker verification system with anti-spoofing capable of filtering synthesized audio in real time. Our stack — ECAPA-TDNN for embedding extraction and CQCC-LCNN for spoof detection.

Problems We Solve

The first is replay attacks: an attacker simply plays a recording. Text-dependent mode is helpless here — anti-spoofing is needed. The second is high voice variability due to colds, fatigue, or noise. Without an adaptive threshold, FRR can exceed 10%. The third is speed: the system must respond in <200 ms, or UX suffers. Moreover, synthesized voices based on WaveNet and Tacotron are becoming increasingly realistic, and traditional methods can't cope.

Attacks on Voice Systems

We distinguish three main types: replay (recording playback), synthesis (WaveNet, Tacotron), and conversion (voice transformation into another). Replay is blocked by adding nonce and timestamp to the request. Synthesis and conversion are detected by CQCC-LCNN trained on ASVspoof 2021 — 98% accuracy at 1% FAR. Replay attack protection reduces losses by up to 90%.

How We Select the Verification Threshold?

The threshold determines the balance between FAR (accepting an impostor) and FRR (rejecting a genuine user). For banking scenarios, FAR <0.5% is needed; for app authorization, 1% is sufficient. We tune the threshold to your scenario using the ROC curve on your data. The table below shows typical thresholds:

Architecture comparison: ECAPA-TDNN gives EER 1.2x lower than x-vectors (0.87% vs 1.05% on VoxCeleb1). For resource-constrained scenarios, we use ResNetSE34L with INT8 quantization — inference on CPU in 50 ms.

Implementation on ECAPA-TDNN

We use a pretrained model from SpeechBrain: ECAPA-TDNN. It outputs embeddings in 192-dimensional space. Inference speed — 80 ms on GPU, 200 ms on CPU. Code:

from speechbrain.pretrained import SpeakerRecognition
import torchaudio

verifier = SpeakerRecognition.from_hparams(
    source="speechbrain/spkrec-ecapa-voxceleb",
    savedir="tmp_verification"
)

def verify_speaker(
    enrollment_audio: str,
    test_audio: str,
    threshold: float = 0.25
) -> tuple[bool, float]:
    """
    enrollment_audio: reference recording of a registered user
    threshold: Accept/Reject threshold (tuned for needed FAR/FRR)
    """
    score, prediction = verifier.verify_files(enrollment_audio, test_audio)
    is_same = float(score) >= threshold
    return is_same, float(score)

Why Anti-Spoofing Is Needed?

Without it, the system is vulnerable: synthesized voice (WaveNet, Tacotron) passes verification. We add an additional classifier based on CQCC-LCNN that distinguishes recordings from live speech. It runs before the main comparison, blocking 98% of attacks. The average project cost with anti-spoofing is $5,000, and monthly savings from fraud prevention reach $15,000.

from speechbrain.pretrained import EncoderClassifier

antispoofing = EncoderClassifier.from_hparams(
    source="speechbrain/asvspoof-cqcc-lcnn",
    savedir="tmp_antispoofing"
)

def is_genuine(audio_path: str) -> bool:
    signal, _ = torchaudio.load(audio_path)
    prediction = antispoofing.classify_batch(signal)
    return prediction[3][0] == "genuine"

Typical Implementation Mistakes

• Collecting only one enrollment phrase is bad. Use 3–5; averaging gives -30% EER.
• Not updating enrollments — voice changes. Re-record every 3–6 months.
• Ignoring replay — add nonce and timestamp to the request.
• Using default threshold — always calibrate on your data.
• Forgetting about noise — minimum SNR 15 dB, otherwise accuracy drops.

Implementation Process

1. Analytics: gather requirements for FAR/FRR, attack types, integration points.
2. Prototype: in 2 days set up the model, test on your recordings, tune threshold.
3. Integration: embed into bot/application via REST API or gRPC.
4. Load testing: verify p99 latency < 300 ms at 500 RPS.
5. Deployment and monitoring: deploy on Kubernetes with autoscaling, log metrics.

What's Included in the Work?

• Documentation on architecture and API specification (OpenAPI).
• Docker image with the model (GPU/CPU version).
• Instructions for deployment and operation.
• Training your team (2–3 days).
• 6-month warranty on the model with possibility of fine-tuning.

We have been in voice biometrics for over 5 years, completed more than 30 projects for fintech and telecom. Average EER in our deployments is 1.2%.

Timelines

Basic system (verification + thresholds) — from 1 week. With anti-spoofing and profile management — 2–3 weeks. Cost depends on the number of instances and load. We estimate projects in 1 day.

Want to test speaker verification on your data? Order a pilot project — we'll adapt the model in 1 day. Get a consultation on your project — we'll send a preliminary estimate.

Speech Recognition and Synthesis: ASR, TTS, Voice Cloning

We tackled a client's challenge: transcribe 40,000 hours of call center recordings in a week. Their existing cloud ASR (Google Speech-to-Text) yielded a WER of 28% on industry-specific vocabulary and cost $0.006 per minute — prohibitively expensive at that volume. The goal was to reduce WER below 10% and switch to self-hosted inference. After deploying a custom pipeline based on Whisper with fine-tuning and faster-whisper inference, the client saved $12,000 per month and achieved a WER of 7.3%.

How does speech recognition ASR handle noisy call center recordings?

The most common issue is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec. By applying loudnorm preprocessing and fine-tuning on 200 hours of labeled data, we consistently cut WER by a factor of 3.

Typical problems we encounter

WER does not converge to the desired metric. Often the culprit is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec.

Diarization fails with more than two speakers. pyannote/speaker-diarization-3.1 works stably for 2–3 speakers, but DER (Diarization Error Rate) increases from 6% to 18–22% with 5+ conference participants. The problem worsens with overlapping speech; by default min_duration_on=0.1 cuts short interjections. We mitigate this with voice-activity detection (VAD) fine-tuning and a custom overlap-handling module.

Voice cloning — latency vs. quality. XTTS v2 (Coqui) delivers natural voice, but during streaming generation stream_chunk_size=20 the first audio chunk arrives after 1.4–2.0 seconds — unacceptable for interactive scenarios. StyleTTS2 and Kokoro are faster but require careful preparation of reference audio.

How do we solve it in practice?

The basic stack for a production pipeline:

• ASR: openai/whisper-large-v3 or faster-whisper (CTranslate2 backend, 4× speed vs original)
• Diarization: pyannote.audio 3.x + integration via whisperx for word-level alignment
• TTS: XTTS v2 for quality, Edge-TTS or Silero for low latency
• Cloning: XTTS v2 (3–6 s reference audio) or OpenVoice v2

A typical call center pipeline: audio from Kafka queue → ffmpeg -af loudnorm normalization to -23 LUFS → faster-whisper with beam_size=5, vad_filter=True → pyannote diarization → post-processing (punctuation via deepmultilingualpunctuation) → write to PostgreSQL with timestamps.

Case study from our practice. A fintech company with 12,000 calls per day. Initial WER on Russian with banking vocabulary — 22% (Google STT). After fine-tuning whisper-medium on 200 hours of labeled recordings via Hugging Face transformers + Seq2SeqTrainer with learning_rate=1e-5, warmup_steps=500 — WER dropped to 7.3%. Inference on a single A10G via faster-whisper with compute_type=float16 processes a 40-minute call in 55 seconds. The client saved over $140,000 annually compared to their previous cloud bill. Contact us for a free pilot estimate to see similar savings on your data.

How to fine-tune Whisper on domain data?

When a general model underperforms, fine-tuning is the first tool. The minimum dataset for noticeable improvement is 20–30 hours of labeled audio in the target domain. Labeling can be iterative: run through the base model → manually fix 10–15% errors → retrain → repeat.

training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=16,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    fp16=True,
    predict_with_generate=True,
    generation_max_length=225,
)

Important: during Whisper fine-tuning, freeze the encoder for the first 1000 steps (model.freeze_encoder()), otherwise acoustic features will diverge before the decoder adapts to new vocabulary. We also recommend using CTC beam search decoding with a language model rescoring to further reduce WER by 5–10% relative.

Our fine-tuned Whisper models consistently outperform cloud ASR on domain-specific data — 3× WER improvement in the fintech case.

Speech synthesis: How to choose a model for your task?

For interactive bots requiring TTFB < 300 ms — Silero or Kokoro. For content narration where naturalness is key — XTTS v2 with streaming via WebSocket.

Our process and deliverables

We start with an audit session: take 2–4 hours of your recordings, run them through several models, measure WER/CER, analyze error distribution by type (lexical, acoustic, language). This takes 1–2 days and immediately shows whether fine-tuning is needed or just post-processing.

Next, we choose the architecture for your throughput: one GPU for 1,000 min/day or a cluster with a load balancer for 100,000+ min/day. Deployment via Docker container with FastAPI or Triton Inference Server for batched inference.

What you get after engagement:

• Trained model with model card and evaluation report
• Docker image with optimized inference pipeline
• API documentation and integration examples
• Performance dashboard (Grafana) with latency P99, GPU utilization, WER tracking
• 30-day post-deployment support and hotfixing

Timelines depend on complexity:

• Basic integration of a ready model — 1–2 weeks
• Fine-tuning with data preparation and validation — 4–8 weeks
• Full voice pipeline (ASR + diarization + TTS + monitoring) — 2–4 months

Project investments typically range from $20,000 to $80,000. Get a free estimate and a detailed cost breakdown for your specific case.

Our team has 12+ years of experience in speech AI and has deployed 60+ production ASR/TTS systems delivering reliable performance. Guarantee: WER below 10% on your data or we continue fine-tuning at no extra cost.

Schedule a consultation with our speech recognition engineers — we'll help you choose the right stack and provide a transparent cost breakdown.