Chapter 07: Cloud API TTS Services Compared

ElevenLabs is the exception: it bills in "credits," where (for most models) 1 character ≈ 1 credit, but the low-latency Flash/Turbo models cost 0.5 credit/char. OpenAI historically quoted per 1M characters for tts-1/tts-1-hd and per minute of audio (token-based) for gpt-4o-mini-tts. We normalize everything to "$ per 1M characters" where possible.

Provider-by-provider

1. ElevenLabs — best naturalness, genuinely good Czech, but the priciest

English: Widely regarded as the most natural-sounding commercial TTS in 2025–2026, especially for expressive long-form reading and audiobooks. Its Multilingual v2 model is the quality reference; Turbo v2.5 trades a little quality for ~250–300 ms latency; Flash v2.5 targets ~75 ms model latency for real-time use.

Czech (cs-CZ): Supported and good. ElevenLabs lists Czech among its 70+ languages (Multilingual v2, Turbo v2.5, and Flash v2.5 all cover Czech), and has a dedicated Czech TTS/voiceover product page. Czech naturalness is among the best available from any cloud API — correct stress and reasonably natural prosody on ordinary prose, clearly better than the older Google WaveNet / Amazon Polly neural Czech voices. It's not flawless (occasional foreign-accent coloring on some cloned/multilingual voices, and numbers/abbreviations can need normalization), but for "natural Czech from a cloud API," ElevenLabs is at or near the top.

Voice cloning: Yes — Instant Voice Cloning (a few seconds of audio, available from paid tiers) and Professional Voice Cloning (30+ min, higher tiers). This is one of ElevenLabs' main differentiators.

Streaming/latency: Yes, WebSocket + HTTP streaming; Flash v2.5 ~75 ms model latency.

Free tier & pricing (verified 2026-05-31):

• Free: ~10,000 credits/month (~10k characters of Multilingual v2, or ~20k via Flash/Turbo at 0.5×), non-commercial, attribution required.
• Starter ~$5/mo: ~30,000 credits, commercial use, instant cloning.
• Creator $22/mo ($11 first month promo historically): ~100,000 credits.
• Pro ~$99/mo: ~500,000 credits.
• Effective cost works out to roughly $0.15–0.30 per 1,000 characters on lower tiers (i.e., $150–300 per 1M chars-equivalent) — much more expensive per character than Google/Azure/Polly. The premium buys naturalness and cloning, not volume economics.

Verdict for the reader: Best Czech quality, but the worst price-per-character of the mainstream options. Good if you read modest volumes and want the best voice; bad if you'll synthesize hours of text and care about "free/near-free."

2. OpenAI TTS (tts-1, tts-1-hd, gpt-4o-mini-tts) — great voices, murky Czech

English: Very natural, especially gpt-4o-mini-tts, which adds steerability — you can pass an instructions field ("speak in a calm, warm tone") to control delivery. Voices: alloy, echo, fable, onyx, nova, shimmer, plus newer ones (ash, coral, sage, ballad, etc.).

Czech (cs-CZ): Unofficial / not a first-class feature. OpenAI does not expose a cs-CZ voice or a language selector — the model auto-detects language from the input text. It can read Czech text and it often sounds passable, but OpenAI does not document Czech as supported, quality is inconsistent run-to-run, and there's an audible English-accent tendency on Czech because the voices are English-trained. Grade: works in a pinch, not reliable for "must sound natural in Czech." Verify with your own Czech samples before depending on it.

Voice cloning: No public custom-voice/cloning in the standard TTS API (the fixed voice set only).

Streaming/latency: Streaming supported (chunked/SSE; realtime via the Realtime API). Latency is good but not the lowest.

Free tier & pricing (verified 2026-05-31):

• No perpetual free tier (new-account credits aside).
• tts-1: ~$15 / 1M characters; tts-1-hd: ~$30 / 1M characters.
• gpt-4o-mini-tts: billed by audio tokens / per minute — roughly $0.015 per minute of audio (≈ a few dollars per million characters equivalent, often cheaper than tts-1-hd). Confirm on the pricing page; OpenAI changed audio billing units in 2025.

Verdict: Excellent English and cheap-ish, but Czech is unsupported officially and accent-prone — fails the reader's hard Czech filter unless your own listening tests say otherwise.

3. Google Cloud Text-to-Speech — the strongest "natural Czech + near-free" pick

English: Multiple tiers. Quality ladder (worst→best): Standard → WaveNet → Neural2 → Studio (premium narration, limited locales) → Chirp 3 HD (2025 flagship, very natural, expressive). There's also Polyglot and instant custom voice on some tiers.

Czech (cs-CZ): Strong and broad — explicitly confirmed. Czech is available across Standard, WaveNet, Neural2, and — critically — Chirp 3 HD. Google's docs explicitly list cs-CZ among the Chirp 3 HD locales (the recently-added batch includes bg-BG, cs-CZ, et-EE, el-GR, hr-HR, hu-HU, lt-LT, lv-LV, ro-RO, sk-SK, sl-SI, sr-RS), served by Chirp 3 HD's 8 named voices (4 male / 4 female) across ~30+ languages. Neural2 Czech is solid; Chirp 3 HD Czech is a clear step up in naturalness. Caveat: for cs-CZ, Chirp 3 HD currently does not support pause control or custom pronunciations (per Google's docs) — fine for plain reading, a limitation if you need SSML-style fine control. This is the most important finding for the reader: Google gives you genuinely modern, natural Czech and a generous free tier.

Voice cloning: Instant Custom Voice is available (consent-gated, enterprise-flavored), but not as frictionless as ElevenLabs. For most readers the prebuilt Chirp 3 HD voices are enough.

Streaming/latency: Streaming synthesis supported (including a low-latency streaming endpoint for Chirp 3 HD); good latency.

Free tier & pricing (verified 2026-05-31, per month, per voice tier):

• Standard: first 4,000,000 chars free/mo, then ~$4 / 1M chars.
• WaveNet / Neural2 / Polyglot: first 1,000,000 chars free/mo, then ~$16 / 1M chars.
• Studio: first 100,000 chars free/mo, then ~$160 / 1M chars (premium).
• Chirp 3 HD: first 1,000,000 chars free/mo, then ~$30 / 1M chars (verify — Chirp pricing has been quoted in the ~$30/1M range and is sometimes grouped with the premium "HD" bucket).

Verdict: Top recommendation for the reader's sweet spot. Natural Czech (Chirp 3 HD), 1M free chars/month on the good tiers, and ~$16–30/1M after that. If you stay under ~1M Czech characters a month, it's effectively free with good-to-excellent quality.

4. Microsoft Azure AI Speech — best breadth of Czech voices + strong free tier

English: Large neural voice catalog, very natural; newer HD / multilingual voices push quality higher and add cross-lingual delivery.

Czech (cs-CZ): Supported with native neural voices — notably cs-CZ-AntoninNeural (male) and cs-CZ-VlastaNeural (female). These are real, native Czech neural voices with good naturalness (comparable to Google Neural2; very usable for reading prose). Azure also has multilingual HD voices that can speak Czech, though native cs-CZ neural voices are the safer bet for accent-free Czech. Grade: solid, reliable native Czech — a strong #2 behind ElevenLabs/Google Chirp 3 HD on quality, but excellent on price.

Voice cloning: Yes — Custom Neural Voice (gated behind an application/approval process for responsible-AI reasons; more enterprise friction than ElevenLabs).

Streaming/latency: Yes, real-time streaming SDK; low latency; good for interactive use.

Free tier & pricing (verified 2026-05-31):

• Free (F0) tier: 0.5 million characters/month of neural TTS at no cost.
• Standard Neural TTS: ~$15 / 1M characters (often quoted ~$16).
• HD voices: higher, roughly $30 / 1M characters (verify — Microsoft has separate HD pricing).
• Custom Neural Voice has additional training/hosting fees.

Verdict: Excellent near-free Czech option. 0.5M free chars/month plus genuine native Czech voices (Antonin/Vlasta). Slightly behind Google on free-tier volume (0.5M vs 1M) and arguably on top-end naturalness, but rock-solid and well-documented.

5. Amazon Polly — cheapest, but Czech is older/standard-only

English: Neural (and 2024+ Generative and Long-Form) voices are good; Generative is the most natural Polly tier. English is competitive.

Czech (cs-CZ): Supported but weak/dated. Polly's Czech voice is Jitka (cs-CZ, female), and historically Czech has been standard-engine only — i.e., the older concatenative/parametric quality, not the natural neural voices. As of mid-2026 Polly's neural/generative coverage for Czech is limited or absent (verify the voice list for your region). Grade: usable for utility readouts, but it does NOT sound as natural as Google/Azure/ElevenLabs Czech. This is the trap to avoid if naturalness is the priority.

Voice cloning: No custom/cloning in Polly (fixed voices only).

Streaming/latency: Yes, supports streaming.

Free tier & pricing (verified 2026-05-31, AWS Free Tier = first 12 months):

• Standard: 5M chars/month free (12 mo), then ~$4 / 1M chars.
• Neural: 1M chars/month free (12 mo), then ~$16 / 1M chars.
• Long-Form: ~$100 / 1M chars.
• Generative: ~$30 / 1M chars (verify).

Verdict: Cheapest mainstream option and great for English, but Czech is standard-tier and noticeably less natural — fails the hard Czech filter for a reading feature where quality matters.

6. PlayHT (play.ht) — flagship is English-first; Czech via legacy voices only

English: Very natural flagship models (Play 3.0 mini, PlayDialog) tuned for conversational/agent use, low latency.

Czech (cs-CZ): Weak. PlayHT markets Czech support and has a "Czech AI voice generator" page, but its best/flagship models are English (and a handful of major languages) only; Czech, where offered, comes from older multilingual voices and is not on par with Google/Azure/ElevenLabs. Treat PlayHT Czech as utility-grade, verify by ear.

Voice cloning: Yes (instant + high-fidelity cloning) — a selling point.

Streaming/latency: Yes, low-latency streaming aimed at voice agents.

Free tier & pricing: Limited free trial; paid plans (Creator ~$31/mo and up) and API pricing roughly in the $10–30/1M-character-equivalent range depending on model/tier. (Pricing structure changes frequently; verify.)

Verdict: Good for English agents and cloning; not a serious natural-Czech option.

7. Deepgram Aura / Aura-2 — fast and cheap, but English-only (and a bit of Spanish)

English: Aura-2 is enterprise TTS optimized for very low latency voice agents and is priced aggressively (~$30/1M chars, sometimes quoted lower). Quality is good for conversational English.

Czech (cs-CZ): Not supported. Aura/Aura-2 TTS covers only ~7 languages — English, Spanish, German, French, Dutch, Italian, Japanese. No Czech TTS. (Note: Deepgram did add Czech to its speech-to-text in late 2025 — don't confuse that with TTS; the read-aloud side has no Czech.) Disqualified for this reader.

Voice cloning: No public cloning.

Verdict: Great for English real-time agents; irrelevant for Czech.

8. Cartesia (Sonic / Sonic-2 / Sonic-3) — ultra-low latency, multilingual, but Czech unconfirmed

English: Sonic is known for extremely low latency (~40–90 ms time-to-first-audio) and strong naturalness — a favorite for real-time voice agents. Sonic-3 (2025) claims state-of-the-art expressiveness.

Czech (cs-CZ): Not clearly supported / unverified. Cartesia advertises a growing multilingual list (English, French, German, Spanish, Portuguese, Chinese, Japanese, and more — on the order of ~15–40 languages depending on model/version). Czech is not clearly on the official supported list as of mid-2026 — if it appears, treat it as new/experimental and test heavily. Grade: assume no reliable Czech until you confirm in the docs.

Voice cloning: Yes (instant voice cloning) — a strength.

Streaming/latency: Best-in-class low latency; great for agents.

Free tier & pricing: Free tier (limited monthly credits) + paid plans; API pricing roughly $15–40 / 1M characters depending on tier/model. Verify.

Verdict: Excellent for English/realtime; Czech is not a confirmed strength — don't bet on it without verifying.

Notable others (brief)

• Speechify, Murf, WellSaid, Resemble.ai: Studio/marketing-oriented. Resemble.ai has strong cloning; Czech support on these ranges from none to limited — verify per product, generally not the price/quality sweet spot for a self-built reading feature.
• iFlytek / Hume / Rime / Inworld: English/Chinese-centric or niche; Czech support weak or absent.

Big comparison table

Czech grades: ✅ native & natural · 🟡 supported but dated/utility-grade · ⚠️ works unofficially / accent-prone · ❌ not supported.

All figures verified 2026-05-31 against vendor pricing/voice pages; some (Chirp 3 HD, Azure HD, Polly Generative, PlayHT, Cartesia) move frequently — re-check before committing.

Which are actually best for Czech?

In honest order of Czech naturalness:

1. ElevenLabs Multilingual v2 — best-sounding Czech from a cloud API, but most expensive.
2. Google Chirp 3 HD (cs-CZ) — very natural and comes with 1M free chars/month → the best quality-per-dollar for Czech.
3. Azure Neural (Antonin / Vlasta) — reliable native Czech, 0.5M free chars/month, great documentation.
4. Google Neural2 (cs-CZ) — solid, also under the 1M free bucket.
5. (distant) Amazon Polly Jitka / PlayHT legacy — utility-grade, audibly dated.
6. OpenAI — might sound fine but unsupported/accent-prone; only after your own Czech listening test.
7. ❌ Deepgram, Cartesia — no/unconfirmed Czech.

For the reader's stated goal (natural Czech + free/near-free): start with Google Chirp 3 HD, with Azure Neural as the backup. Both keep you at $0 under ~0.5–1M Czech characters/month, which is a lot of reading. Reserve ElevenLabs for the texts where you want the very best voice and can accept the per-character cost.

Minimal API snippets for the top 3

Google Cloud TTS — Chirp 3 HD, Czech (Python):

# pip install google-cloud-texttospeech ; set GOOGLE_APPLICATION_CREDENTIALS
from google.cloud import texttospeech as tts

client = tts.TextToSpeechClient()
resp = client.synthesize_speech(
    input=tts.SynthesisInput(text="Dobrý den, toto je test českého hlasu."),
    voice=tts.VoiceSelectionParams(
        language_code="cs-CZ",
        name="cs-CZ-Chirp3-HD-Charon",  # verify exact Chirp 3 HD cs-CZ voice name in docs
    ),
    audio_config=tts.AudioConfig(audio_encoding=tts.AudioEncoding.MP3),
)
open("out.mp3", "wb").write(resp.audio_content)

Azure AI Speech — Czech neural voice (Python):

# pip install azure-cognitiveservices-speech
import azure.cognitiveservices.speech as speechsdk

cfg = speechsdk.SpeechConfig(subscription="YOUR_KEY", region="westeurope")
cfg.speech_synthesis_voice_name = "cs-CZ-VlastaNeural"  # or cs-CZ-AntoninNeural
synth = speechsdk.SpeechSynthesizer(speech_config=cfg)
synth.speak_text_async("Dobrý den, toto je test českého hlasu.").get()

ElevenLabs — Multilingual v2, Czech (Python):

# pip install elevenlabs
from elevenlabs.client import ElevenLabs
from elevenlabs import save

client = ElevenLabs(api_key="YOUR_KEY")
audio = client.text_to_speech.convert(
    voice_id="VOICE_ID",                 # pick a multilingual voice from your library
    model_id="eleven_multilingual_v2",   # or eleven_turbo_v2_5 / eleven_flash_v2_5 (0.5 credit/char)
    text="Dobrý den, toto je test českého hlasu.",
    output_format="mp3_44100_128",
)
save(audio, "out.mp3")

Key Takeaways

• Most "best" TTS APIs are English-first. Deepgram Aura-2 (English-only), Cartesia Sonic (Czech unconfirmed), and PlayHT's flagship models effectively fail the Czech filter — don't be seduced by their English demos.
• For natural Czech + free/near-free, Google Cloud TTS (Chirp 3 HD) is the top pick: native, modern, very natural Czech with 1M free characters/month and ~$16–30/1M after.
• Azure Neural (Antonin / Vlasta) is the strong runner-up: reliable native Czech, 0.5M free chars/month, excellent docs/SDK.
• ElevenLabs has the best-sounding Czech of any cloud API but is by far the most expensive per character — use it selectively, not for bulk.
• Amazon Polly's Czech (Jitka) is standard-engine/dated — cheap and fine for utility readouts, not for a quality-sensitive reading feature.
• OpenAI sounds great in English and is steerable, but Czech is unofficial and accent-prone — verify by ear before relying on it.
• Billing is per input character (≈$ per 1M chars); 1M chars ≈ ~18–22 hours of audio, so the free tiers cover a lot of reading.
• Re-verify pricing and Czech voice availability at call time — TTS tiers, free quotas, and new-model language coverage (especially Chirp 3 HD and Cartesia) changed repeatedly across 2025–2026.

Key References

• ElevenLabs. ElevenLabs API Pricing. 2026. https://elevenlabs.io/pricing/api — Credit-based pricing tiers, models (Multilingual v2, Turbo v2.5, Flash v2.5), free-tier credits.
• ElevenLabs. Czech Text to Speech (cs-CZ). 2026. https://elevenlabs.io/text-to-speech/czech-cs-cz — Confirms Czech support and Czech voice availability across models.
• Cybernews. ElevenLabs Pricing 2026: Costs, Tiers, Hidden Limits & Honest Review. 2026. https://cybernews.com/ai-tools/elevenlabs-pricing/ — Independent breakdown of tiers, credits, commercial-use limits.
• Google Cloud. Supported voices and languages — Cloud Text-to-Speech. 2026. https://cloud.google.com/text-to-speech/docs/list-voices-and-types — Czech (cs-CZ) voices across Standard/WaveNet/Neural2/Chirp 3 HD.
• Google Cloud. Chirp 3: HD voices. 2026. https://cloud.google.com/text-to-speech/docs/chirp3-hd — Chirp 3 HD locale coverage incl. Czech, instant custom voice.
• Google Cloud. Text-to-Speech pricing. 2026. https://cloud.google.com/text-to-speech/pricing — Per-tier free monthly character allowances and per-1M-char pricing.
• Microsoft. Language and voice support — Azure AI Speech. 2026. https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support — Czech neural voices (AntoninNeural, VlastaNeural) and HD/multilingual coverage.
• Microsoft Azure. Speech Services pricing. 2026. https://azure.microsoft.com/en-us/pricing/details/cognitive-services/speech-services/ — Free F0 tier (0.5M chars/mo), Neural and HD per-1M-char pricing.
• Amazon Web Services. Supported voices in Amazon Polly (Voice List). 2026. https://docs.aws.amazon.com/polly/latest/dg/voicelist.html — Czech voice Jitka and engine (standard) availability.
• Amazon Web Services. Amazon Polly Pricing. 2026. https://aws.amazon.com/polly/pricing/ — Standard/Neural/Long-Form/Generative pricing and free-tier characters.
• OpenAI. Audio and speech guide / API pricing. 2026. https://platform.openai.com/docs/guides/audio and https://openai.com/api/pricing/ — tts-1/tts-1-hd/gpt-4o-mini-tts models, voices, steerability, pricing.
• Deepgram. TTS Models / Models & Languages Overview. 2026. https://developers.deepgram.com/docs/tts-models — Aura/Aura-2 language coverage (English-focused; no Czech).
• Cartesia. Sonic Text to Speech. 2026. https://cartesia.ai/sonic — Sonic/Sonic-3 latency, multilingual list, cloning (Czech not clearly listed).
• PlayHT. Best Czech AI Voice Generators in 2025. 2025. https://play.ht/blog/best-czech-ai-voice-generators/ — Comparative view of Czech TTS quality across providers.