TTS Model Quality Ranking 2026: How to Compare Models

A useful TTS model ranking starts with the recording you need to produce. A voice that makes a convincing short announcement may stumble over a chapter with abbreviations, quotations, and long pauses. A model that wins a public listening comparison may require infrastructure you cannot operate. Quality, deployment, rights, and cost deserve separate decisions.

Correction, September 30, 2026: The previous version combined approximate scores from incompatible sources. We removed that numeric table and its unsupported winner claims. We do not have an archived benchmark export that reproduces those rows. This article provides a comparison procedure rather than a current official leaderboard or a claim that we ran a controlled listener study.

What a Quality Score Actually Answers

The Artificial Analysis Speech Arena is a starting point for investigating listener preferences. Read the source’s current methodology, filters, and model/provider labels before quoting results. Save the source and the date together. A model family name is insufficient when different providers, voices, versions, or output settings can appear under similar names.

An aggregate preference result does not establish pronunciation accuracy for your vocabulary, suitability for your audience, or reliability on your device. A small difference in a score needs context from uncertainty and participation. It also cannot tell you whether a browser implementation uses the same weights, precision, text normalization, or voice that appeared in the benchmark.

Do not translate parameter count into a quality rank. Kokoro’s upstream card identifies an 82-million-parameter model; that is an architecture fact. Browser download size depends on quantization, runtime files, voice data, and caching. Neither number is a measured listening score. Similarly, a voice catalog letter grade is an editorial discovery aid rather than a standardized MOS result.

Shortlist by Deployment Requirements

OfflineTTS integrates five engines: Kokoro, Kitten, Piper, Supertonic, and Pocket. Start with the workspace that matches your constraints, then audition its output. These are workflow distinctions, not an ordered quality ranking.

Candidate	Reason to audition it	Boundary to check
Kokoro	Preset voices for narration and pronunciation review	Non-English Kokoro text uses our phonemization endpoint before local synthesis
Kitten	Compact English model and expression presets	Presets are fixed options; they do not guarantee directed emotion
Piper	CPU-oriented generation with curated speakers	Voice-specific model licenses and pronunciation vary
Supertonic	Multilingual browser generation	Validate language tags, selected voice, and names in the script
Pocket	Language-bundle selection and an authorized reference recording workflow	Check cloning consent, reference quality, and device resources
Hosted speech service	Managed APIs and production features	Review the selected model, contract, network dependency, and current price

A cached local engine can reduce dependence on a remote inference service. It still consumes memory, battery, storage, and review time. The surrounding website downloads assets and may load services according to your choices. See our privacy policy for the implemented data paths rather than assuming “local” describes every request.

Build a Representative Test Script

Use material from the actual project that you have permission to process. For a tutorial, include menu labels, product names, version numbers, and a code identifier. For an audiobook, include dialogue and a paragraph that crosses a sentence boundary. For accessibility, include headings, dates, punctuation, and link text. Keep sensitive information out of external benchmark submissions.

Create three samples: a short introduction, a difficult sentence, and a longer passage. One example of a difficult sentence is: “On 3 September, Dr. Lee reviewed version 2.4, the A/B results, and a twelve-percent change.” Decide the intended pronunciation of the date, abbreviation, slash, and percentage before listening. This gives you a correction checklist instead of a vague impression.

Normalize the input consistently. If one engine receives expanded abbreviations and another receives raw text, the comparison includes your preparation work. That may be appropriate for evaluating the whole production workflow, but record the difference. Preserve both the source script and the exact submitted version so a later model update can be evaluated against the same input.

Reproducible Model Audition Record

Record the engine and version, provider, voice ID, language, speed, and relevant generation settings. For local inference, record browser, operating system, available memory, acceleration backend, and whether assets were cached. For hosted inference, record the network conditions and any queue or rate-limit response. Separate initial download time from a warmed generation.

Keep an output file and a short note for each render. Useful fields include time to first playable audio, total render time, output duration, failed segments, pronunciation corrections, pause edits, and artifacts. Do not publish a universal speed multiplier from a single laptop. Repeating a warm test helps identify variation, but it does not represent every supported device.

For a listening comparison, conceal engine labels and alternate playback order. Use the same playback level without processing one candidate more aggressively. Ask reviewers about intelligibility, phrasing, fatigue, and suitability for the script. Record how many people participated and their language familiarity. A small internal audition is useful; describe it as an internal audition, not a scientific blind study.

Keep Cost Separate from Preference

Local synthesis has no OfflineTTS subscription or per-character charge. It does not have a zero total operating cost. A production estimate should include hardware, energy, storage, model download traffic, maintenance, and correction time. If you host an upstream model yourself, include deployment, monitoring, updates, and access controls as well.

A hosted provider’s bill may use characters, audio duration, tokens, credits, or plan allowances. Check the official pricing page on the date you decide, including discounts, overage, taxes, and commercial-use terms. Compare the price of completed, reviewed audio rather than assuming one credit equals one character across products. If no current quote is available, mark cost as unverified instead of inserting an approximate number.

Verify Rights before Publishing

Treat application code, runtime, model weights, preset voices, reference recordings, and source text as separate assets. An Apache or MIT software license does not automatically settle every person’s voice rights or grant permission to narrate copyrighted material. Redistribution obligations can differ from simply using an output file.

Kokoro’s upstream model card identifies Apache 2.0 for its weights. Read the license conditions, including applicable notices when redistributing licensed material. Piper voice files may have their own model-card licenses. Pocket reference recordings require permission from the speaker and any recording rights holder. Consult the model source and license information and the site terms for the selected workflow.

Do not infer that a paid provider has “no privacy” or that a local engine has an unconditional privacy guarantee. Hosted services have specific policies and contractual options; local applications have their own network and storage boundaries. Evaluate the actual content path and retention policy for the project.

Decide with the Finished Output

Choose a candidate that clears pronunciation, rights, device, and data-handling requirements first. Then compare the editing effort and listening experience of the remaining candidates. If two are close, a more reliable export or a simpler correction process can matter more than a public preference score.

Save the record with the final audio. Repeat the difficult sentence after changing voices, upgrading a model, or altering normalization settings. Long productions should include a chapter-sized review before a full batch. This prevents a successful ten-second demo from becoming an unsupported promise about a whole book.

Open the browser workspace to audition supported engines, or use our Arena interpretation guide when reading an external leaderboard. Neither page claims a universal quality winner.