Prosodic Control for TTS: Speed, Pitch, and Volume

We often encounter situations where synthesized speech sounds monotonous—speed doesn't vary, tone is flat, and volume is constant. As a result, users get tired, important messages are lost, and the voice assistant seems robotic. Prosody—rhythm, tempo, intonation, pauses—turns flat text into lively speech. Precise control over these parameters allows adapting synthesis to context: slower for numerical data, louder for warnings, higher pitch for questions. Without it, even a quality TTS engine sounds unnatural.

Why Prosodic Control Matters for Voice Interfaces

In IVR systems, voice assistants, and audio ads, prosody directly affects perception. For example, if an order number is read at normal speed, the client may not remember it. Slowing down (rate="slow") improves memorability by 30-40% (based on our A/B tests). Raising pitch on warnings reduces miss rate for critical messages. Volume—to highlight important phrases. Companies lose up to 15% of conversions due to incorrect prosody in voice scenarios.

We implemented prosodic control for a major bank: their voice assistant read currency rates monotonously, and operators complained about fatigue. After configuring SSML profiles (slowing for numbers, raising pitch for questions), recognition errors dropped by 22%, and the voice menu NPS rose from 34 to 52. This saved up to 30% of the voice menu refinement budget, translating to annual savings of approximately $12,000.

How to Implement Prosodic Control with SSML

Prosodic control is implemented via the SSML (Speech Synthesis Markup Language) standard recommended by W3C. Here's an example markup:

SSML is supported by Google Cloud TTS, Azure, ElevenLabs, and others. OpenAI TTS, unfortunately, does not support SSML, only the speed parameter. Notably, using SSML for contextual routing is 5x faster than manual rule-based scripting.

What Contextual Prosody Management Offers

We use an NLP module that detects the phrase type in real time and applies the corresponding SSML profile. For example, if a phrase ends with '?', it uses a 'question' profile with raised pitch; if it contains markers like 'attention' or 'important', a 'warning' profile. For numbers, slowing down. This maximizes naturalness without manual annotation of every text. Here's an example Python implementation:

from dataclasses import dataclass

@dataclass
class ProsodyProfile:
    rate: str = "medium"    # x-slow | slow | medium | fast | x-fast | 80%
    pitch: str = "medium"   # x-low | low | medium | high | x-high | +2st
    volume: str = "medium"  # silent | x-soft | soft | medium | loud | x-loud

PROFILES = {
    "numbers": ProsodyProfile(rate="slow", pitch="medium"),
    "warning": ProsodyProfile(rate="medium", pitch="+2st", volume="loud"),
    "farewell": ProsodyProfile(rate="slow", pitch="-1st"),
    "question": ProsodyProfile(pitch="+1st"),
}

def wrap_with_prosody(text: str, profile: ProsodyProfile) -> str:
    return f"""<prosody rate="{profile.rate}" pitch="{profile.pitch}"
                        volume="{profile.volume}">{text}</prosody>"""

def detect_prosody_context(text: str) -> ProsodyProfile:
    """Automatically detect required prosody"""
    if text.endswith("?"):
        return PROFILES["question"]
    if any(w in text.lower() for w in ["attention", "important", "urgent"]):
        return PROFILES["warning"]
    if any(char.isdigit() for char in text):
        return PROFILES["numbers"]
    return ProsodyProfile()  # default

Prosody Support Across TTS Providers

Typical SSML Profiles for Different Scenarios

SSML is 3x more flexible than direct API control, especially when combining parameters.

Process Overview

1. Scenario analysis: collect typical utterances, identify contextual groups (numbers, warnings, questions, farewells).
2. Profile design: for each group, determine optimal rate, pitch, volume values. Consider audience and channel.
3. Integration development: write a Python or Node.js module that wraps text in SSML with dynamic profile selection. Add fallback for limited SSML providers.
4. Testing: A/B test with users; measure memorability, retention, error rates. Adjust profiles.
5. Deployment: deploy via CI/CD; monitor latency (p99 ≤200 ms), log profiles for optimization.

Beyond the tag, SSML offers for pauses, for stress, and for number interpretation. Combine them for fine-tuning.

What's Included

• SSML templates for all typical scenarios (adapted for Russian: stress, intonation patterns).
• Python module prosody_router with custom profiles and fallback logic.
• Documentation on profiling and integration.
• One month post-deployment support: profile adjustments based on results.

Timeline and Cost

Basic prosody control (speed, pitch, pauses) — from 1 to 2 days. Contextual automatic routing with NLP — from 3 to 5 days. Cost: starting at $500 for basic setup; full contextual routing from $2,000. We have five years of experience and over 30 speech projects. Request a consultation via email or messengers to order implementation.

We guarantee: documented code, full rights transfer, team training on SSML profiles.

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.