Speech Synthesis on AWS: Integrating Amazon Polly with Python and SSML

When integrating a voice assistant on AWS, many clients face a common problem: standard Amazon Polly speech synthesis for Russian sounds unnatural. Neural TTS for ru-RU is not supported, and using Western accents is not an option. Additionally, there is a text length limit — when synthesizing via synthesize_speech, you can pass a maximum of 1500 characters. We solve these challenges using SSML markup, asynchronous processing via Lambda and S3, and custom prosody settings. Over 5 years of work we have implemented more than 50 TTS cloud projects, each with its own nuances.

How to bypass the Neural limitation for Russian speech?

The main tool is SSML (Speech Synthesis Markup Language). It allows you to control pauses, tempo, stress, and even pronunciation of individual characters. For example, if you need digits to be read individually, use <say-as interpret-as="digits">. To create natural pauses, use <break time="300ms"/>. Here is a production-ready markup example we use:

<speak>
  Здравствуйте! Ваш заказ <break time="300ms"/>
  номер <say-as interpret-as="digits">12345</say-as> готов.
  <prosody rate="slow">Пожалуйста, проверьте данные.</prosody>
</speak>

This brings the sound closer to human-like even on standard voices. According to AWS documentation, SSML corrects intonation better than simply changing voice parameters. Two voices are available for Russian:

For comparison, Azure Neural TTS supports Russian, but its cost is roughly 2–3 times higher for the same volume. Amazon Polly with SSML delivers 80% of Azure's quality at a lower price.

Why choose integration via Lambda and S3?

We use the Lambda + S3 combination for scalable and cost-effective synchronization. The client sends text — the Lambda function synthesizes speech via boto3 and saves the file to S3. The user receives a Presigned URL for direct download. For long texts, we launch an asynchronous start_speech_synthesis_task — this saves resources and avoids Lambda's time limits.

import boto3

polly = boto3.client('polly', region_name='us-east-1')

def synthesize_speech(text: str) -> bytes:
    response = polly.synthesize_speech(
        Text=text,
        OutputFormat='mp3',      # mp3 | ogg_vorbis | pcm | json
        VoiceId='Tatyana',       # Maxim | Tatyana for ru-RU
        LanguageCode='ru-RU',
        Engine='standard',       # standard | neural (not for ru-RU)
        SampleRate='22050',      # 8000 | 16000 | 22050
        TextType='text',         # text | ssml
    )
    return response['AudioStream'].read()

# SSML synthesis
ssml_text = """
<speak>
  Здравствуйте! Ваш заказ <break time="300ms"/>
  номер <say-as interpret-as="digits">12345</say-as> готов.
</speak>
"""
response = polly.synthesize_speech(
    Text=ssml_text,
    TextType='ssml',
    OutputFormat='mp3',
    VoiceId='Tatyana',
)

For long texts:

# For long texts — async task to S3
response = polly.start_speech_synthesis_task(
    Text=long_text,
    OutputFormat='mp3',
    VoiceId='Tatyana',
    OutputS3BucketName='my-tts-bucket',
    OutputS3KeyPrefix='audio/'
)
task_id = response['SynthesisTask']['TaskId']

How SSML helps in practice: a case with educational video lectures

In one project — voicing educational video lectures — we used custom SSML templates to improve number and formula intelligibility by 30% without using Neural voices. We configured pronunciation for special characters: <say-as interpret-as="digits"> for numbers, <phoneme alphabet="ipa" ph="pi">π</phoneme> for Greek letters, and <prosody rate="85%"> for slow reading of complex terms. This kept costs low (standard synthesis) while achieving high perceptual quality.

What is SSML: key elements for speech synthesis

These tags help overcome standard voice limitations and make speech more natural.

What's included in Amazon Polly integration work

• Complete integration code with boto3, including SSML templates
• Documentation on architecture and API call
• Cost optimization recommendations based on your volume
• Test synthesis on your data for quality assessment
• Team training (1 hour online) on SSML basics and working with Polly
• 2 weeks of post-release support for prompt issue resolution

Work process

1. Analysis: we examine your scenario — text volume, required languages, quality requirements.
2. Architecture: we design the scheme — Polly + S3 + Lambda + CloudFront (optional).
3. Implementation: we write Python code with boto3, configure SSML for your texts.
4. Testing: we run on real data, measure p99 latency.
5. Deployment: we deploy to your AWS account, grant access.

Timeline: from 2 to 5 business days depending on complexity. Cost is calculated individually.

Guaranteed results and support

We have been integrating AWS and TTS for over 5 years, all engineers are certified. We provide full documentation and ready-made scripts so you can maintain the system yourself. We train your team on Polly and SSML, assist with debugging during testing. After project completion, you are left with a working solution and a clear understanding of its architecture.

Want to evaluate synthesis quality on your own data? Contact us — we will prepare a demo sample and calculate the optimal configuration for your scenario. Order a consultation to discuss details.

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.