Self-Hosted Speech Synthesis with Voice Cloning Using Coqui TTS

Self-Hosted Speech Synthesis with Voice Cloning

We often encounter situations where a client needs high-quality Russian speech synthesis, but using cloud APIs (Google Cloud TTS, Amazon Polly) isn't an option—data leaves the premises, and monthly bills for thousands of minutes can blow a startup's budget. Coqui TTS solves both problems: it's an open-source library that can be deployed on your own servers and supports fine-tuning for any voice.

At a typical load of 100,000 characters per month, self-hosted Coqui TTS saves up to $500 compared to Google Cloud TTS. We've integrated Coqui TTS into production for several fintech projects (IVR, voice assistants) and accumulated expertise in model selection, inference tuning, and latency optimization. This article covers how to quickly set up TTS on your own hardware, which models actually work for Russian, and how to achieve quality indistinguishable from a human voice.

Clients often come with a task to build a voice assistant in CRM or an IVR system. Typical requirements: the voice must sound natural, support pauses and intonation, and handle specific terminology. Cloud APIs either lack the desired voice in Russian or become expensive at high volumes. We offer an alternative—Coqui TTS on your GPU.

Voice Cloning Mechanism in XTTS v2

One of Coqui's key features is voice cloning from a reference audio. The XTTS v2 model takes a short recording (3–10 seconds) and synthesizes speech with the same timbre. We use this approach to generate voices for virtual assistants—just one minute of a speaker's speech is enough for the model to reproduce intonations and mannerisms.

from TTS.api import TTS

# Initialize XTTS v2
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")

# Synthesize in Russian
tts.tts_to_file(
    text="Hello! This is an example of speech synthesis in Russian.",
    speaker_wav="reference_speaker.wav",  # reference voice (3–10 sec)
    language="ru",
    file_path="output.wav"
)

# Streaming synthesis (chunks)
for chunk in tts.tts_with_vc_streaming(
    text="Long text for streaming synthesis",
    speaker_wav="reference.wav",
    language="ru"
):
    # process audio chunk
    pass

Why Coqui TTS Is Better Than Cloud Services

Comparison of key characteristics:

This comparison shows that for high-load projects or strict privacy requirements, self-hosted TTS is the only reasonable choice.

Which Models Are Suitable for Russian?

Out of the box, Coqui TTS supports Russian in VITS and XTTS v2 models. VITS ru is a lightweight model for basic synthesis, XTTS v2 is multilingual with cloning. We recommend XTTS v2 for production: quality is close to commercial solutions, and speed is sufficient for real-time.

tts = TTS("tts_models/ru/cv/vits")  # Russian VITS model
tts.tts_to_file(
    text="Hello world",
    file_path="output.wav"
)

How We Integrate Coqui TTS into Your Project

Our approach goes beyond simply installing the library. We conduct an audit, select the model for your load (up to 100 requests/sec? need Triton?), optimize latency via batch inference and FP16.

Work process:

1. Analysis — voice requirements, language, load, use case (IVR, podcasts, assistant).
2. Model selection — XTTS v2, VITS, or fine-tuned for the client.
3. Integration — FastAPI wrapper, Kubernetes deployment, monitoring.
4. Fine-tuning — if needed, improve diction, remove artifacts.
5. Testing — latency p99 measurements, MOS quality assessment.
6. Deployment — into your infrastructure or our managed server.

What's Included

• Ready-made FastAPI wrapper with /tts and /clone endpoints.
• Docker container for GPU deployment.
• API documentation (OpenAPI spec).
• Performance testing scripts.
• GPU selection recommendations (from RTX 3060 to H100).
• 1 month of support after delivery.

We guarantee synthesis will work with latency <200 ms (p99) under single-threaded inference (XTTS v2 on RTX 3080). Basic integration takes 2 to 5 days depending on complexity. If fine-tuning is required, an additional 1–2 days for computations.

Based on our experience (over 20 TTS projects, 5 years in the AI/ML market), we find the optimal balance between quality and speed. Contact us for a project estimate—we will assess your task free of charge and propose an architecture solution.

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.