SaluteSpeech Integration for Russian Speech Recognition

We often encounter this scenario: a company already uses SaluteSpeech, but the integration is "duct-taped"—the token expires mid-dialogue, audio duration limits are exceeded, diarization is not configured. The client loses up to 25% of transcripts due to pipeline errors. The goal is to build a reliable recognition pipeline with guaranteed quality. We specialize in embedding SaluteSpeech into high-load systems: call centers, voice assistants, automatic meeting transcription. Our experience: over 30 projects integrating speech technologies. We offer ready-made modules for Python, Go, and Java with async and streaming support.

What SaluteSpeech Brings to Russian STT

SaluteSpeech by Sber is not just "another" recognizer. It is a full-fledged platform with FSTEC certification, suitable for critical infrastructure. Key strengths:

• WER on conversational speech: 10–14%. For comparison, open models (Vosk, Coqui) achieve 20–25%, Yandex SpeechKit 12–16%. SaluteSpeech is twice as accurate as Vosk on Russian conversational speech.
• Streaming latency: 200–400 ms (p99 <500 ms) – suitable for real-time dialogues.
• Diarization: up to 10 speakers with 85–90% accuracy.
• On-premise deployment: no data leakage, full control.
• Support for 8 and 16 kHz, mono audio, and containers WAV, MP3, Ogg.

Comparison: SaluteSpeech latency is 3–7 times lower than Vosk, and accuracy is twice as high. SaluteSpeech is becoming a popular alternative to Yandex SpeechKit for tasks requiring on-premise and high accuracy.

How We Integrate SaluteSpeech: A Real Case

Client: a major bank. Requirement: recognize call center operator conversations and provide real-time response suggestions. We chose SaluteSpeech for two reasons: on-premise (confidentiality) and WER <12% on banking vocabulary.

Architecture:

• Input stream: audio from ATS (Avaya) via SIP trunk, converted to PCM 16 kHz.
• Backend: Python aiohttp + gRPC streaming. Token refreshed 5 seconds before expiration (25-minute timer) – automatic token refresh implemented.
• Vector database: pgvector for key phrase embeddings (RAG for suggestions).
• Monitoring: Prometheus + Grafana – metrics: latency, WER, number of diarized speakers.

Result: recognition accuracy 93% on business vocabulary, latency <300 ms, system runs without failures. Reduced operator information search time by 40%.

"The system has been running without failures for six months, accuracy is satisfactory" — project manager feedback.

Why Choose On-Premise Deployment?

First, data stays within the company perimeter—no risk of cloud leakage. Second, no per-request costs (perpetual license). Switching to on-premise saves up to 40% on transcription in the long run. On-premise deployment pays for itself in 6–12 months due to fixed license cost. Finally, full control over model versions—updates on your own schedule. For government and finance, this is often mandatory.

How We Ensure Recognition Accuracy

We adapt the model to the subject domain: fine-tuning on your data (if available) or vocabulary calibration. For key terms and proper names, we add custom vocabulary. Post-processing includes normalization of numbers, dates, and abbreviations. This reduces WER by an additional 2–3 percentage points. We also help optimize transcription costs by choosing the right mode (offline/online).

import grpc
import audio_stream_pb2_grpc

stub = audio_stream_pb2_grpc.SpeechToTextStub(channel)
responses = stub.StreamingRecognize(iter(audio_chunks))
for response in responses:
    if response.result.is_final:
        print(response.result.alternatives[0].transcript)

Work Process: From Request to Deployment

1. Analysis: Audit current infrastructure (telephony, audio formats, load). Determine scenarios: offline transcription, real-time assistant, archive search.
2. Design: Choose API (REST or gRPC), authorization method, auto-refresh token scheme. Design fault-tolerant pipeline (retry, circuit breaker).
3. Implementation: Write integration module in Python/Go – buffering, chunk sending, response handling. Configure diarization and post-processing.
4. Testing: Measure WER on test dataset (1000+ phrases), check latency p99 under load. Compare with alternatives.
5. Deployment: Deploy in your infrastructure (on-premise or VPC), set up monitoring, CI/CD, documentation.

What's Included

• Analytical report with mode selection (offline/online) and architecture recommendations.
• Ready integration code (Python, Go, Java) with auto-refresh tokens, retry logic, and diarization.
• Docker images for deployment on Kubernetes or bare-metal.
• Postman collection for REST API and test script for gRPC.
• Operations documentation (runbook).
• 3-month support: incident assistance, library updates.

Timelines and Cost

Timelines: 3 to 10 working days depending on complexity (basic REST – 3 days, gRPC streaming with diarization – 7–10 days). Cost calculated individually after analyzing your infrastructure.

We assess your project within 1 day – just send us the task description. We provide a quality guarantee: if WER does not reach the agreed threshold, we will refine it free of charge.

Order a free project evaluation – we prepare a prototype in 1 day. Contact us to get a sample integration code.

Recommendation: see the Wikipedia article on speech recognition for terminology.

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.