Custom voice: TTS model fine-tuning with MOS 4.3+ guarantee

Every third response from your voice assistant sounds unnatural—trembling timbre, missing phonemes. We solve this by fine-tuning a TTS model on the client's voice. After fine-tuning on 30–60 minutes of recordings, the model steadily reads any text: MOS rises to 4.3+ from 3.8 in zero-shot, and reverse recognition WER drops by 5–10%. Result: the assistant stops 'stuttering' even on complex queries.

Why fine-tuning over zero-shot?

Zero-shot cloning (e.g., XTTSv2 in speaker encoder mode) gives acceptable results but suffers from timbre trembling, artifacts on rare phonemes, and instability on long texts. Fine-tuning on 30–60 minutes of the target voice locks in the speaker's acoustic space, reduces reverse recognition WER by 5–10%, and increases UTMOS by 0.3–0.5. Main advantages: predictable quality on any input, ability to augment data (noise, reverberation), and control over intonation via conditioning.

What goes into dataset preparation for TTS fine-tuning?

Minimum volume: 30 minutes of clean recordings. Optimal: 1–2 hours. Audio requirements: sampling rate 22050 or 24000 Hz, signal level –18…–12 dBFS, signal-to-noise ratio >30 dB, clip lengths 3–15 seconds.

Preparation steps:

• Record in a studio or quiet room (check for background noise).
• Clean noise: use HPSS filter or spectral subtraction.
• Sentence-level segmentation: force alignment with Montreal Forced Aligner.
• Validate duration and quality: run through a validation script.

import pandas as pd
from pathlib import Path
import soundfile as sf
import numpy as np

def validate_dataset(dataset_dir: str) -> dict:
    """Check dataset before training"""
    metadata = pd.read_csv(f"{dataset_dir}/metadata.csv",
                           sep="|", names=["file", "text"])
    stats = {
        "total_files": len(metadata),
        "total_duration": 0,
        "errors": []
    }

    for _, row in metadata.iterrows():
        wav_path = f"{dataset_dir}/wavs/{row['file']}.wav"
        if not Path(wav_path).exists():
            stats["errors"].append(f"Missing: {wav_path}")
            continue

        audio, sr = sf.read(wav_path)
        duration = len(audio) / sr
        stats["total_duration"] += duration

        if sr != 22050:
            stats["errors"].append(f"Wrong SR {sr}: {wav_path}")
        if duration < 1.0 or duration > 15.0:
            stats["errors"].append(f"Bad duration {duration:.1f}s: {wav_path}")

    stats["total_duration_min"] = stats["total_duration"] / 60
    return stats

Fine-tuning XTTS v2—stack and configuration

We use the official Coqui TTS repository with modifications for commercial tasks. Below is the config for fine-tuning only the decoder (faster, less noise).

from trainer import Trainer, TrainerArgs
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts

config = XttsConfig()
config.load_json("base_xtts_config.json")

# Fine-tuning parameters
config.audio.output_sample_rate = 24000
config.batch_size = 4
config.eval_batch_size = 2
config.num_loader_workers = 4

# Fine-tuning only decoder (faster, less data)
config.trainer_args = {
    "epochs": 100,
    "save_step": 1000,
    "print_step": 50,
    "eval_split_size": 0.1
}

Variations: you can fine-tune the entire encoder+decoder if dataset >2 hours, but this increases training time 2–3x and requires caution with overfitting.

How to evaluate synthesized voice quality?

The primary metric is MOS (Mean Opinion Score) per ITU-T P.800. We use an internal panel of 10–15 listeners, each evaluating 50–80 samples. Results:

Objective metrics:

• UTMOS: automatic naturalness score (MOS-predictor model)
• SECS (Speaker Embedding Cosine Similarity): similarity to donor voice >0.95
• WER on reverse recognition: no more than 5% at medium pace

Infrastructure and training cost

GPU selection depends on budget and required speed. We recommend configurations with minimal FLOPS:

Training cost depends on the chosen configuration and data volume. Savings compared to buying a ready-made TTS solution can reach 30–50%. We help select a configuration within your budget.

What's included in our TTS fine-tuning project?

1. Source material audit—evaluate recording quality, noise, diction.
2. Dataset preparation—cleaning, volume normalization, segmentation (force alignment).
3. Model training—choose architecture (XTTS, IhreTTS, YourTTS), tune hyperparameters.
4. Quality evaluation—MOS, UTMOS, SECS, WER.
5. Model export—ONNX / TorchScript for inference.
6. Integration—API wrapper, testing in your product.
7. Documentation and team training—how to update the voice, extend fine-tuning.

We guarantee: final MOS at least 4.0 with a dataset from 30 minutes. If not met, we redo at our cost.

Estimated timelines

How to avoid common fine-tuning pitfalls

Recordings with background noise are the main enemy of quality. We apply HPSS filter and VAD segmentation. Phoneme imbalance (e.g., missing unvoiced or sibilant sounds) is compensated by a specialized script to create a balanced dataset. On small data (<30 minutes), L2 regularization and early stopping help. All these measures ensure stable results without overfitting.

If you have questions about dataset, architecture, or budget—contact us for a consultation. Request a cost estimate for your project—we will find the optimal solution for your needs.

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.