Best Text to Speech APIs (2026): Top 10 Compared

ElevenLabs sells its speech models to developers under the ElevenAPI name, alongside separate creative and agent products. This entry covers the text to speech API.

The models page lists four text to speech options. Eleven v3 is described as the most expressive model, with 70 or more languages, support for multi-speaker dialogue and a 5,000 character limit per request. Eleven v3 Conversational is a real-time variant with latency of about 280 ms and audio tags for fine control of delivery. Eleven Multilingual v2 covers 29 languages, allows 10,000 characters per request and is described as the most stable choice for long-form generation. Eleven Flash v2.5 is the low-latency option at about 75 ms, covering 32 languages with a 40,000 character limit.

Pricing is published per 1,000 characters. Flash and Turbo cost US$0.05 and Multilingual v2 and v3 cost US$0.10, and the same unit rates apply across plans. The plans differ in the characters included each month: the free tier includes 20,000 Flash characters, Starter is US$6 per month, Creator US$22 (first month US$11), Pro US$99, Scale US$299 and Business US$990. Enterprise adds custom DPA and SLA terms, SSO and BAAs for HIPAA customers.

The same account also reaches speech to text, voice changing, voice isolation, sound effects, music and dubbing APIs. A startup grants programme offers 12 months free with 33 million characters.

ElevenLabs fits developers for whom voice quality and language breadth matter more than the lowest price per character.

Deepgram is best known for speech recognition, and its Aura models add text to speech to the same platform. This entry covers the text to speech API. Deepgram's voice agent API, which bundles recognition, a language model and synthesis, is compared separately on the AI voice agent platforms page.

The documentation shows a simple request pattern: text is posted to the speak endpoint and audio streams back. Voices are selected by model name, such as aura-2-thalia-en. REST and WebSocket interfaces are both available, and SDKs are published for JavaScript, Python, Go, .NET and Java.

Language coverage is narrower than several competitors. The documentation lists English with American, British, Australian, Irish and Filipino accents, Spanish with several regional accents, German, French, Dutch, Italian and Japanese.

Pricing is published per 1,000 characters. On pay as you go, Aura-2 costs US$0.030, Aura-1 costs US$0.015 and the newer Flux TTS model costs US$0.045. The Growth plan lowers those to US$0.027, US$0.0135 and US$0.0405 in exchange for pre-paid annual credits starting at about US$4,000. New accounts get US$200 in free credit, and the vendor states there are no minimums and no credit expiry. At the time of checking, the pricing page also advertised a promotion matching Flux TTS spend with credits up to US$500 until the end of 2026.

The pricing page lists SOC 2 Type 1 and Type 2 certification. Deepgram fits teams that already use or plan to use its speech recognition and want one vendor for both directions of a real-time audio pipeline.

Cartesia builds the Sonic family of text to speech models and also sells speech recognition and managed voice agents. Its documentation describes Sonic 3.5 as the company's fastest and most natural model, with latency under 90 ms and native support for 42 languages, including several Indian languages such as Hindi, Tamil, Telugu, Bengali and Marathi. The pricing page lists Sonic-3.6 in its plan comparison.

The documentation emphasises details that matter in customer-facing speech. It states that Sonic reads order numbers, phone numbers, IDs and email addresses naturally without preprocessing, and that English heteronyms such as read and bass are pronounced correctly from context.

Versioning is unusually explicit. A dated model ID never changes. The plain sonic-3.5 ID follows the latest stable snapshot, and sonic-latest may change without notice and is not recommended for production.

Plans are credit-based. Free includes 20,000 credits a month, or about 27 minutes of speech. Pro at US$5 includes 100,000 credits, about 133 minutes, and adds a commercial licence and instant voice cloning. Startup at US$49 includes 1.25 million credits, about 1,667 minutes, and professional voice cloning. Scale at US$299 includes 8 million credits, about 10,667 minutes, with priority support. Enterprise adds custom concurrency, DPAs and BAAs, SSO and a shared Slack channel.

Cartesia fits developers who put latency first. Buyers expecting high parallel volume should confirm concurrency, since self-serve plans allow between 2 and 15 simultaneous text to speech requests.

Murf started as a voiceover studio for creators and now sells a separate developer API. The API offers two speech models. Falcon is built for conversational AI and real-time voice agents, and the pricing page states time to first audio under 130 ms. Gen 2 is aimed at studio-quality narration for multimedia.

The Falcon 2 documentation lists use cases including customer support agents, collections reminders, healthcare assistants, lead qualification and IVR. It states time to first audio under 100 ms, support for more than 10,000 concurrent calls at enterprise scale, 150 or more voices in 20 or more styles, and a multinative capability that lets a voice switch languages within a sentence.

Regional deployment is a notable feature. The documentation lists endpoints in the US, Canada, India, South Korea, the UAE, Japan, Australia, the EU, the UK and Sao Paulo. Enterprise customers can use isolated regional environments for data residency. Default concurrency is 15 on US-East and 2 on other regional endpoints, with higher limits on request.

Pricing is flat. Falcon costs US$0.01 per 1,000 characters and Gen 2 costs US$0.03, with voice changing at US$0.10 per minute and translation at US$0.02 per 1,000 characters. A free trial provides US$10 in credit each month with 5 concurrent requests, pay as you go credit does not expire, and a startup programme offers US$1,500 in credits over three months.

Murf fits cost-sensitive, high-volume conversational deployments and teams that need speech processed in a particular region.

Inworld began in AI characters for games and now sells its speech models as a general developer API. Its documentation describes Realtime TTS models with context-aware synthesis, zero data retention and voice cloning, available through streaming and non-streaming APIs.

Two models are offered. Realtime TTS-2 is the flagship, with natural language steering to shape delivery, support for more than 200 languages and locales, instant voice cloning and timestamps that include phonetic details and visemes. The vendor gives its time to first audio byte as 100 ms. Realtime TTS-2 Flash is the lower-cost option, documented at 20 ms, with the same language list and instant cloning. The latency figures are stated as P90 values measured server-side, excluding network latency.

Pricing is published per million characters. On-demand use costs US$25 for TTS-2 and US$15 for Flash, with up to 70 minutes included, 100 custom voices and a commercial licence. Monthly plans include credits and reduce the rate. Creator at US$25 lowers TTS-2 to US$20 and Flash to US$10. Builder at US$100 lowers them to US$17.50 and US$9. Developer at US$300 lowers them to US$15 and US$8. Growth at US$1,500 lowers them to US$12.50 and US$7 and adds ZDR, HIPAA and BAA add-ons. Enterprise is quoted as low as US$5 per million for TTS-2, with EU and India data residency and price matching.

The same credits can be spent on speech recognition and on more than 220 language models through a router. Inworld fits developers who want low per-character cost at volume.

Hume AI is a research company focused on emotionally expressive voice AI. It offers two APIs: the Empathic Voice Interface for real-time speech-to-speech conversation, and Octave for text to speech.

The documentation describes Octave as a text to speech system built on language model intelligence. Named use cases include narration for video, podcasts and audiobooks, lessons with varied emotional tone, and voices for digital avatars and game characters. Octave 2 is in preview with expanded language support and lower latency.

Voices come from three sources. A library offers more than 100 voices designed by Hume. Voice design creates a new voice from a descriptive prompt. Voice cloning builds a voice from a recorded or uploaded sample with the speaker's consent. SDKs are published for TypeScript, Python and React.

Pricing is by subscription with included characters. Free includes 10,000 characters, about 10 minutes. Starter at US$3 includes 30,000. Creator at US$14 (first month US$7) includes 140,000. Pro at US$70 includes 1 million. Scale at US$200 includes 3.3 million and Business at US$500 includes 10 million. Overage is charged from Creator upward at US$0.15, US$0.12, US$0.10 and US$0.05 per 1,000 characters as the plan rises. Requests per minute range from 15 to 225, and concurrent connections from 1 to 30.

Hume fits developers whose product depends on emotional range, such as narration and character-driven experiences. High-throughput transactional workloads should check the request-per-minute caps and compliance terms, which are fullest on Enterprise.

OpenAI offers text to speech through the speech endpoint of its Audio API. The guide describes three models. The newest, gpt-4o-mini-tts, is recommended for intelligent real-time applications and can be prompted to control accent, emotional range, intonation, impressions, speed, tone and whispering. The older tts-1 model has lower latency, and tts-1-hd produces higher quality at higher cost.

A request takes three inputs: the model, the text and a voice, with an optional instruction such as asking for a cheerful and positive tone. Output defaults to MP3 and other formats can be selected, and responses can be streamed for real-time playback. The endpoint provides 13 built-in voices, with marin and cedar recommended for the best quality. The older models support a subset of nine voices. OpenAI states that the voices are optimised for English, and its usage policies require a clear disclosure to end users that the voice they hear is AI-generated.

Pricing depends on the model. gpt-4o-mini-tts is billed by tokens, at US$0.60 per million text input tokens and US$12 per million audio output tokens. tts-1 costs US$15 per million characters and tts-1-hd costs US$30 per million characters. Its cost per minute therefore depends on the length of the generated audio.

The main case for OpenAI is consolidation. A team already using OpenAI language models can add speech output with the same SDK, key and invoice, and use written instructions rather than markup to shape delivery. The absence of custom or cloned voices and the English focus are the trade-offs.

Google Cloud Text-to-Speech converts plain text or Speech Synthesis Markup Language input into audio. The documentation includes quickstarts for the command line and client libraries, a reference for supported voices and languages, SSML reference material and guides to device profiles for different playback hardware.

The service spans several generations of voice technology, each priced separately. Standard and WaveNet voices cost US$4 per million characters after 4 million free characters a month. Neural2 and the Polyglot preview voices cost US$16 per million after 1 million free. Chirp 3 HD voices, which Google groups under its latest models, cost US$30 per million after 1 million free. Studio voices cost US$160 per million after 1 million free. Instant custom voice costs US$60 per million characters with no free tier.

Gemini-TTS is billed differently. Gemini 2.5 Flash TTS and Flash-Lite preview cost US$0.50 per million text input tokens and US$10 per million audio output tokens. Gemini 2.5 Pro TTS and the Gemini 3.1 Flash TTS preview cost US$1 and US$20. The pricing page states that audio tokens accrue at 25 per second of audio.

Billing counts every character sent, including spaces, newlines and most SSML tags. Billing must be enabled on the project, and charges begin once the free allowance is exceeded.

Google fits organisations already on Google Cloud that want one provider for bulk, low-cost speech and premium voices. Buyers should compare tiers carefully, since the gap between Standard and Studio pricing is forty-fold.

Amazon Polly is the AWS service for converting text into speech. The documentation describes four voice types: generative, long-form, neural and standard. Neural voices also support a Newscaster speaking style for news narration.

Two features set Polly apart. Speech Marks are metadata returned alongside or instead of audio, marking when each word or sound occurs, which is used to synchronise an avatar's lips or highlight text as it is read. Polly also allows generated speech to be cached and replayed at no additional cost, which suits apps that repeat the same prompts.

Pricing is per million characters processed, for either speech or Speech Marks requests. Standard voices cost US$4, Neural voices US$16, Generative voices US$30 and Long-Form voices US$100. The free tier includes 5 million Standard characters a month, plus, for the first 12 months only, 1 million Neural, 500,000 Long-Form and 100,000 Generative characters a month. New AWS customers since July 2025 may also receive up to US$200 in general Free Tier credits.

The pricing page gives worked examples. A typical three-page news article of about 6,500 characters costs US$0.03 with Standard voices and US$0.20 with Generative voices. A conversational app with 300 short responses per user per month costs US$0.12 to US$3.00 per user depending on voice type. AWS GovCloud pricing is listed at US$4.80 for Standard and US$19.20 for Neural.

AWS documents Polly as certified for HIPAA and PCI DSS workloads. Polly fits teams already on AWS, particularly in regulated sectors, education and accessibility.

Azure AI Speech is Microsoft's speech service, now presented as part of Foundry Tools. Its text to speech capability offers standard neural voices out of the box and custom voice options for brands. Microsoft's Voice Live API for real-time agents is a separate product.

The documentation states that standard voices cover more than 100 languages and locales, with each voice model available at 24 kHz and high-fidelity 48 kHz. Synthesis runs in real time through the Speech SDK or REST API. For content longer than 10 minutes, a batch synthesis API processes requests asynchronously and returns files to download.

Custom voice has two forms. Professional voice fine-tuning trains a unique brand voice, and personal voice creates a voice from a short sample. Microsoft states that custom voice is a limited access feature and requires an application.

The Azure pricing page shows a free tier with 0.5 million characters a month for neural voices. Its price table loads in the browser, so rates were read from Microsoft's public retail price list for the East US region. Pay as you go rates there are US$15 per million characters for neural voices, US$22 for Neural HD, US$100 for long audio, US$24 for custom neural real-time synthesis and US$24 for personal voice. Custom voice training is listed at US$52 per hour and model hosting at about US$4.03 per hour. Commitment tiers start at 80 million characters for US$960 a month. Microsoft states that prices are estimates and vary by agreement, region and currency.

Azure fits enterprises already standardised on Microsoft that need wide language coverage, long-form batch output and a governed route to a custom brand voice.