best 16 ai text-to-speech models

elevenlabs spans the whole range — flash and turbo v2.5 for low-latency agent use, multilingual v2 and the expressive v3 for narration — across 30+ languages, with instant and professional voice cloning, websocket streaming, and a conversational-ai stack for real-time agents. the sdks, voice library and documentation are best-in-class, which is a real reason it's the safe default.

the trade-offs: it's the priciest per character among the majors, and flash's ~75ms latency is a vendor claim with no independent benchmark. if you want quality with zero infrastructure, it's the tool; if you're cost-sensitive at high volume, weigh the open models below.

cartesia's sonic line is built for real-time: streaming-first architecture, instant voice cloning, and low latency by design, billed at one credit per character with a genuine free tier and plans scaling from $5/mo. for building a voice agent where time-to-first-audio is the metric that matters, it's the most natural pick in the category.

the honest caveats: the credit accounting is more opaque than a flat per-character rate, its sub-100ms claim is unbenchmarked, and while quality is very good it isn't always the arena leader. for agents specifically, that's usually the right trade.

openai offers tts-1 and tts-1-hd at flat per-character rates and gpt-4o-mini-tts, a token-priced model you can steer with natural-language 'instructions' for tone and emotion — something tts-1 can't do. it ships around eleven voices with streaming, and the killer feature is integration: if you're already on the openai sdk it's one client and one auth.

the limits are clear: no voice cloning, fewer voices than the specialists, english-centric though multilingual works, and token-priced audio makes cost forecasting imprecise. for existing openai builders who want cheap, steerable speech, it's an easy add.

deepgram's aura-2 is a tts model aimed squarely at english voice agents — a large voice set, low-latency streaming, and clean rest/websocket apis, with $200 in free credit for new accounts. its real advantage is pairing with deepgram's speech-to-text so your whole agent loop (listen and speak) comes from one vendor with one integration.

it's english-focused with narrower language coverage than google or azure, its sub-250ms latency claim is unbenchmarked, and it isn't a cloning or narration tool. for english ivr and voice agents where stt+tts unity matters, it's a strong, simple choice.

google spans a huge catalogue: cloud text-to-speech from $4/million characters (standard/wavenet) up through neural2, chirp 3 hd and studio, plus the newer prompt-controllable, multi-speaker gemini tts (token-priced), across 40+ languages with ssml and a large free tier. for breadth and a low price floor, nothing here beats it.

the friction is the sprawl — the multi-tier catalogue is genuinely confusing to price — and some gemini tts models carry preview labels, so confirm general availability before you commit. cloning is limited (custom voice is a separate gated program). for gcp-native shops it's an obvious pick.

kokoro is an 82-million-parameter model small enough to run on cpu, with several english and some multilingual voices, and the thing that sets it apart from the higher-quality open models is the licence: both the code and the kokoro-82m weights are apache-2.0, so it's genuinely commercial-safe. self-host it for effectively nothing, or use it on fal.ai at $0.02 per 1k characters.

it uses fixed voice packs, so there's no zero-shot cloning, and its expressiveness sits below the big proprietary models. its repo commit cadence has slowed (last substantive push in 2025), worth watching for staleness. for high-volume, non-cloning narration where cost matters most, it's the pick.

azure ai speech offers neural tts across 140+ languages and locales with the richest ssml controls (styles, roles, prosody), streaming, and a gated custom neural voice program, at $16/million characters (neural hd $22, reduced in march 2026) with commitment tiers lower and a 500k/month free tier. for enterprises that need compliance, data regions and fine ssml control, it's the safe corporate pick.

the trade-off is weight: setup is heavier than the newer api-first tools, and custom voice requires approval. if you're already on azure and need governance, it fits naturally.

rime targets telephony and voice agents with mist v2 for high volume and the expressive arcana flagship, offering pronunciation control, low latency, 3,000 free minutes to start, and — unusually — on-prem and vpc deployment for teams that can't send audio to a public api. for contact-center use where getting names and numbers right matters, that reliability focus is the draw.

it's a smaller brand with a narrower voice library than the leaders, and its on-prem latency figures are vendor claims. for high-volume agent workloads needing pronunciation reliability and deployment control, it earns its spot.

chatterbox, from resemble ai, does zero-shot voice cloning from about ten seconds of audio, with exaggeration/emotion control and built-in perth watermarking, and — the differentiator versus f5 and fish — both its code and weights are mit-licensed, so it's genuinely commercial-safe. it's very actively maintained and available self-hosted or via hosts like fal.ai and wiro. a multilingual variant exists.

resemble self-reports that listeners prefer it over elevenlabs, which is a vendor a/b test, not an independent result — check tts arena v2 instead. self-hosting means gpu ops. as the open cloning model with a clean licence, it's the standout.

polly is aws's steady tts: 30+ languages, ssml, speech marks, and a newer generative expressive tier, priced from $4/million characters (neural $16, generative $30) with a 12-month free tier and up to $200 in new-account credit. if you live in aws and want cheap, reliable speech that just works, it's the obvious pick.

its top-line expressiveness trails the specialist leaders and there's no zero-shot cloning. that's the trade for stability and price — for many production apps, the right one.

hume's octave was the first tts to take natural-language emotional prompts rather than ssml — you tell it how a line should be delivered, in words — with expressive, steerable output, voice design and multilingual support. for characters, companions and expressive narration where emotional delivery is the point, it does something the others don't.

the caveats: pricing tiers are reported inconsistently across third parties (confirm on hume's own page), and its ecosystem is smaller than the majors. if emotional control is your feature, it's worth the look.

cosyvoice, from alibaba's funaudiollm, is multilingual (cjk plus english and more) with streaming, zero-shot voice cloning, and cross-lingual and instruction-following synthesis in its 2/3 line — and crucially it's apache-2.0, so it's commercial-safe. for multilingual self-hosting, especially where asian languages matter, it's the best open option.

the trade-offs are practical: the documentation is research-grade and it's heavier to stand up than a tiny model like kokoro. alibaba also offers hosted variants via dashscope if you'd rather not run it yourself (verify that pricing separately).

fish speech / openaudio s1 delivers some of the best open-model quality here — zero-shot cloning, multilingual, low latency — and it's a common reason developers reach for open tts. but the trap is the licence: the code is apache-2.0 while the openaudio s1 weights are cc-by-nc-sa-4.0, non-commercial and share-alike, so you cannot ship the open weights in a commercial product.

the commercial path is fish audio's paid api (per-character/credit on fish.audio — verify the current rate), not the downloadable weights. great for research and demos self-hosted; for production, budget for the api. this is the single place developers most often get burned in this category.

orpheus, from canopy labs, is a llama-3.2-3b-based tts with eight voices, emotive tags (laugh, sigh and the like), and low-latency streaming, all under apache-2.0 so it's commercial-safe. self-host it or use it on fal.ai at $0.05 per 1k characters. for an expressive, permissively-licensed open model aimed at agents, it's a good pick.

it's english-centric and, being a 3b model, needs a real gpu for real-time use, so its ~200ms streaming figure assumes proper hardware and is a vendor claim. within those bounds it's one of the more expressive open options.

resemble ai offers high-quality voice cloning, real-time synthesis and emotion control, billed unusually per second (about $1.80 per hour of audio; $0.001/sec on the voice-agent path), with a distinctive companion product line for deepfake detection and watermarking (resemble detect). it's also the company that open-sourced chatterbox, which speaks to its cloning expertise.

the friction is commercial: per-second billing is hard to compare against per-character rivals, and voice clones are monthly add-on fees. for a product that needs cloning plus a security/authenticity narrative, it's a coherent choice.

voxcpm, from openbmb, is a compact (~0.5b) tokenizer-free model with context-aware zero-shot cloning and expressive output, apache-2.0 and therefore commercial-safe, with a realtime voxcpm2 variant and a hosted listing on wiro at $0.0001/char. it's the newest of the high-star open expressive models and an attractive option for developers who want something current and permissive.

being young, it has fewer production references than kokoro or cosyvoice, and its quality standing should be checked on tts arena v2 rather than asserted. as a fresh, permissive, expressive open model, it rounds out the list.