50 featured voice models across 26 providers. A selection of what Onepin routes to, planned and validated through the same pipeline.
50 models
Alibaba's real-time Qwen-Audio 3.0 tier: first audio in about 300ms across 16 languages, directed in plain language ('slower, warmer') or with inline tags like [giggles].
Plain-language instruction (emotion, role, pace), rate, pitch and volume, plus inline emotion tags.
The quality tier of Qwen-Audio 3.0, tuned for naturalness and timbre fidelity over speed. Same 16 languages and plain-language direction as Flash, and #1 on the Artificial Analysis TTS leaderboard at launch.
Plain-language instruction (emotion, role, pace), rate, pitch and volume, plus inline emotion tags.
Alibaba's real-time Qwen-Audio 3.1 tier, with first audio in about 300ms. Its multilingual voices cover Mandarin, Chinese dialects and nine more languages, alongside British and American English voices.
Inline tags in the script for emotion and non-verbal sounds, like [excited] or [laughing].
Amazon's long-form TTS for audiobooks and podcasts, with custom lexicons and full SSML phoneme control across 30+ languages.
Designed for long-form audio (audiobooks, podcasts) with full SSML.
Boson's open-weights model, served hosted: 100+ languages detected automatically, zero-shot cloning from a single reference clip, and inline tags for emotion, style and sound effects.
Inline tags for emotion, style (whisper, shout, sing), sound effects, pace and pauses. No numeric speed or pitch, no SSML.
Massively multilingual TTS (140+ languages) with cross-lingual voice transfer and reference-audio emotion for global dubbing.
Reference-audio emotion, speed, age and gender, plus cross-lingual voice transfer.
Cartesia's May-2026 flagship: #1 on the Artificial Analysis Speech Arena with ~82ms latency, 50+ emotions and IPA dictionaries.
Speed and volume, 50+ emotions, <break> and <spell> tags and pronunciation dictionaries.
Cartesia's August-2026 stable release, preferred over Sonic 3.5 nearly two to one in blind listening: context-aware pacing without tags, native alphanumerics, and 44 languages on the same voices.
Speed and volume, 60 emotion values (beta), <break> and <spell> tags, [laughter], and IPA or sounds-like pronunciation dictionaries.
Cartesia's rolling beta channel: it always points at the next Sonic build ahead of the stable release, so you hear upcoming changes early. It can change without notice and is meant for testing, not production.
Same levers as the current Sonic release (speed, volume, emotion, <break>, <spell>, pronunciation dictionaries), subject to change between builds.
Sub-200ms conversational English TTS built for voice agents, with SSML phoneme control and a natural, on-brand delivery.
Speed, pitch and volume via SSML, emphasis tags and a conversational style.
English TTS for voice agents that holds dialogue context across turns and reports what was already spoken when a caller cuts in. First audio lands in ~80ms.
Speed 0.5 to 1.5 in 0.05 steps and an expressivity dial (beta), with no SSML or pause tags.
Lowest-latency ElevenLabs model (~75ms) for real-time agents, with strong Chinese and Cantonese and 300+ voices.
Stability, similarity and speed (limited expression).
Studio-grade multilingual speech with wide locale coverage and full SSML. Best-in-class Chinese and Cantonese.
Stability, similarity, style exaggeration and speed (0.7 to 1.2).
Balanced quality and speed with audio markup tags and instant cloning; sub-250ms latency for interactive use.
Stability, similarity, style and speed.
Alpha flagship with inline audio-tag emotion ([whispering], [laughing]) across 70+ languages. The most expressive ElevenLabs model.
Audio tags ([whispering], [laughing], [excited]) plus stability, similarity and style exaggeration.
ElevenLabs' newest flagship, built on v3 with closer voice matching and sharper audio-tag delivery. 90+ languages, and a voice reads each one with a native accent rather than its source accent.
Audio tags ([whispering], [shouting], [laughing]), inline IPA in /slashes/, stability and similarity. No speed control or SSML.
Expressive successor to S1 with open-ended [bracket] emotion tags, 64+ expressions and ~100ms latency; strong CJK coverage.
Open-ended [bracket] emotion tags, 64+ expressions and paralinguistics (laugh, sigh, whisper).
Fish Audio's S2.1 successor with open-ended [bracket] emotion tags, word-level emphasis and strong coverage across 13 locales.
Open-ended [bracket] emotion cues, [emphasis] and paralinguistics such as breaths, laughs and pauses.
Gemini-native TTS with natural-language style prompts and 200+ expressive audio tags across single- and multi-speaker output.
Natural-language style prompts and 200+ expressive audio tags ([whispers], [laughs]).
DeepMind WaveNet voices across 40+ languages with full SSML control. The workhorse for high-volume, multilingual pipelines.
SSML prosody (rate/pitch/volume), emphasis, breaks, say-as and phoneme tags.
Google's flagship Gemini TTS and the successor to 3.1 Flash. It detects 130+ input languages on its own and voices them from a library of regional-accent personas.
Inline tags for vocal events and pauses (<laugh>, <sigh>, <short pause>) and CAPS for emphasis. No SSML, IPA, or numeric speed and pitch.
The high-throughput tier of Gemini 3.8 TTS, built for volume work and real-time agents. Same voice library as Flash, with 100+ languages detected automatically.
Inline tags for vocal events and pauses (<laugh>, <sigh>, <short pause>) and CAPS for emphasis. No SSML, IPA, or numeric speed and pitch.
Gradium's TTS for voice agents: median first audio under 250ms, instant cloning from a 10-second clip and word-level timestamps for tight sync. English, French, German, Spanish and Portuguese.
Speed, randomness (temperature) and voice similarity, <break> tags up to 2 seconds, and a pronunciation dictionary.
An 82M-parameter open-weight model from hexgrad, released under Apache-2.0. Each preset voice is pinned to one language. English holds over half of them and reads the most naturally.
Speed from 0.25x to 4x. No SSML, pause tags, emotion control, or IPA.
A 0.5B open-weight speech model from inclusionAI, released under Apache-2.0. Its 4 built-in voices read Mandarin Chinese and English.
No speed, SSML, pause tags, emotion control, or IPA.
The Max tier of Inworld's 1.5 line, with inline IPA control and expressive non-verbal markers for real-time agents.
Inline IPA, non-verbal markers and emphasis (full NL steering on TTS-2).
Inworld's research-preview flagship (May 2026), #1 on the Artificial Analysis Speech Arena, with LLM-style natural-language steering.
Natural-language steering in [brackets], three stability modes and audio-context conditioning.
Inworld's fastest and most cost-efficient TTS-2 tier, with about 20ms server-side time to first byte and 200+ languages.
Break pauses, inline IPA pronunciation and non-verbal tags such as [laugh] and [sigh]; no natural-language steering.
A 40M-parameter open-weight model from KittenML, released under Apache-2.0, that runs on CPU. Its 8 preset voices read US English only.
Speed from 0.5x to 2x. No SSML, pause tags, emotion control, or IPA.
An 80M-parameter open-weight model from KittenML, released under Apache-2.0, that runs on CPU. Its 8 preset voices read US English only.
Speed from 0.5x to 2x. No SSML, pause tags, emotion control, or IPA.
A 15M-parameter open-weight model from KittenML, released under Apache-2.0, that runs on CPU. Its 8 preset voices read US English only.
Speed from 0.5x to 2x. No SSML, pause tags, emotion control, or IPA.
Azure's HD tier: natural, conversational delivery that adapts tone in context, with full SSML and custom lexicon support.
Natural conversational tone with speed and full SSML.
Azure's neural voices with the broadest language coverage in the market: full SSML, emotional styles and custom lexicons.
Full SSML: prosody, emotional styles (cheerful/sad/angry) and viseme output.
Microsoft's newest MAI flagship TTS (Build 2026 preview) with expressive styles, multi-speaker and long-form via Azure Speech SSML.
Expressive styles with style degree, 15+ emotions, multi-speaker and long-form.
Microsoft's low-latency MAI voice model: about twice the speed of MAI-Voice-2 across 18 locales, built for voice agents and interactive lines.
Per-line emotion and delivery style through SSML express-as.
Microsoft's highest-fidelity MAI voice model: 23 languages, one voice identity across all of them, and long-form narration that holds speaker consistency.
Per-line emotion and delivery style through SSML express-as, multi-speaker dialogue, and long-form generation.
Multilingual HD flagship (Jan 2026) with deep controllability: emotion, voice mixing, pinyin/Jyutping tones and IPA dictionaries.
Emotion, speed/pitch/volume, voice mixing, pause and interjection tags, plus IPA dictionaries.
Low-latency multilingual voice (Jul 2026) with the same emotion, voice mixing and pronunciation controls as Speech 2.8 HD at a lower price.
Emotion, speed/pitch/volume, voice mixing, pause and interjection tags, plus pronunciation dictionaries.
Studio-friendly voices for content teams, with a pronunciation library, IPA overrides and simple speed, pitch and emphasis controls.
Speed, pitch, emphasis, pause and pronunciation overrides.
Korea's market-leading TTS, with premium Korean voices plus English, Japanese and Chinese, avatar video and cloning.
Emotion (where supported), speed, pitch, volume and pause.
A 357M-parameter open-weight model from NVIDIA, released under the NVIDIA Open Model License. Its 5 preset voices keep one speaker across English, Korean, German, and French. All five are native English speakers, so English reads the most naturally.
No speed, SSML, pause tags, emotion control, or IPA.
Steerable OpenAI voices driven by a natural-language instructions field ('speak slowly and warmly'). Simple, fast and inexpensive.
Natural-language instructions field ('speak slowly and warmly') plus speed.
A 1.7B-parameter open-weight model from Qwen, released under Apache-2.0. Nine preset voices speak all ten languages: English, Chinese, Japanese, Korean, German, French, Russian, Spanish, Italian, and Portuguese. Each voice reads most naturally in its native language.
No SSML, pause tags, emotion control, or IPA.
Studio-grade, ethically-sourced voice cloning used by EA, Sony and Lucasfilm; speech-to-speech that captures a real performance.
Pitch, emotion and formant driven by a reference performance (speech-to-speech), plus de-noising.
Licensed voices from real actors on Respeecher's Voice Marketplace, each with its own narration styles, and the talent is paid on every use. Rendered per order, not streamed: built for narration and dubbing, not live agents.
Per-voice narration style and language. No speed, pitch or SSML.
A sub-100ms naturalness leader (Coval bench) tuned for effortless, hands-off conversational speech, with deliberately minimal controls.
Minimal by design: playback speed and voice selection only.
Low-latency conversational model with 300+ voices and deterministic IVR pronunciation for high-volume agents.
Speed, deterministic IVR pronunciation and pause control.
Rime's enterprise conversational model with time to first byte under 100ms and an engine rebuilt for concurrent load, on the same voices as Mist v2 across English, Spanish, German and French.
Speed (whole line or per bracketed word), custom pauses in brackets and spell() for letter-by-letter reading. No inline pronunciation.
Speechify's multilingual Simba model for English, German, Spanish, French, Italian, and Brazilian Portuguese. Each line is voiced in its locale's language by a native-speaker voice.
<break> tags for pauses. No speed, pitch, emotion, or IPA control.
Speechify's recommended English model and the most expressive in the Simba family. Voices cover US, UK, Indian, Australian, and Nigerian English.
<break> tags for pauses. No speed, pitch, emotion, or IPA control.
Don’t see your model?
Turn any script into production-quality voice,
in any language, in minutes.
