Compare local TTS models: languages, hardware and speed evidence

Decide whether you need narration, multilingual announcements or a voice resembling a particular speaker. Then choose a language and actual test passage. Kokoro lists eight languages, Qwen3-TTS ten and Chatterbox Multilingual V3 23+; those counts are not pronunciation-quality rankings. Listen to the same short passage with names, numbers and punctuation.

Even when your language is listed, accent or terminology may not meet your needs. Broad language coverage is different from accurate reading of a particular regional or professional script. Track “listed as supported” separately from “passed my sample.”

Kokoro is an 82M option for ordinary narration; its official README provides CPU/CUDA examples and Apple Silicon M1–M4 MPS fallback. Qwen3-TTS separates CustomVoice, VoiceDesign and Base workflows; its official quick start describes Python 3.12 and CUDA. The official quick start reviewed here does not document Qwen MPS or CPU⁴ paths.

Chatterbox Multilingual V3 is a 500M multilingual model; its repository's Python example selects CPU, CUDA or MPS. Those options document execution paths, not equal performance. MPS documentation can make it a candidate for Mac users, but speed and memory still need local testing.

Qwen's README says first audio may arrive in as little as 97 ms, but gives no device, input length, precision, batch or completion boundary. This is a streaming TTFA claim. Separate author-run benchmark repositories publish RTX 4090/5090 results for Qwen-derived CUDA paths; read them with the runtime and conditions shown below.

These are neither site measurements nor results from the official Qwen model team. The faster-qwen3-tts author reports CUDA-graph RTF⁵ 4.78/4.22 and TTFA 156/174 ms for 0.6B/1.7B on an RTX 4090, but the table omits input length and repetitions.

The Triton repository's RTX 5090 results specify BF16⁶, batch 1, two Korean/English texts, three warmups and 20 measured runs. Both repositories define RTF as actual audio duration divided by generation time, the reciprocal of the conventional RTF defined here; larger values mean faster-than-real-time output.

Both sets describe particular optimized stacks.

The Triton table separately reports initial load time: 6.0 seconds for Hybrid and 17.5 seconds for PyTorch⁷ Base. It does not give the actual output duration of the two test passages, so do not transfer these RTF and request times to text of another length.

A shared example clip in a demo is not necessarily generated by the selected model and is not a quality comparison. To compare output, generate the same text and speaker conditions with each model and verify the source and settings.

If the finished file is A seconds long and the measured interval is T seconds, RTF is T divided by A. If elapsed time equals audio duration, RTF is 1. The reciprocal, A divided by T, is the real-time multiplier, also 1 in that example. Using the same label for both can reverse interpretation.

For interactive speech, TTFA may matter; for long-form delivery, use the final-chunk completion time. Record whether cold load, preprocessing, synthesis, vocoder/decoder, chunk joining and file writing are included. For batches, distinguish one request's latency from total throughput⁸.

Within the supported scope of each model, fix one language, passage, target duration and speaker type. Log exact model ID/revision, OS, CPU/GPU, Python and library versions, backend, precision, batch size⁹ and sample rate¹⁰. Run once from a fresh process, separate load from generation, then repeat the same task after warm-up.

Measure the saved WAV's actual duration and calculate RTF for the synthesis interval. Keep cold and warm runs¹¹ separate and retain every run, not just the mean or fastest. Log clipping, pronunciation errors and unnatural pauses separately from speed. A speed ranking cannot stand in for quality review.

If the official docs offer no path for your OS, change the candidate hardware or model rather than guessing performance. MPS support does not imply CUDA-equivalent speed. If one path is slow, inspect logs and device use: CPU fallback, unsupported operations, audio writing or memory pressure may be involved.

The official evidence does not establish a universal speed or memory winner among these models. Once a required language, workflow and documented acceleration path are confirmed, your own RTF and quality notes can inform a purchase. Leaving an unknown field blank is safer than guessing.

Kokoro's official model card lists Apache-2.0, and the Chatterbox repository lists MIT. Qwen's repository Apache-2.0 notice covers code; check the card for the exact checkpoint¹². For distribution or commercial use, review terms for code, weights, input audio and generated speech separately.

Use reference audio only from yourself or a speaker who has authorized it. When trying a demo, start with a supplied voice; if it requires uploading reference audio, check how the file is stored and processed. Before publishing generated speech, consider whether it could be mistaken for a real person's statement and make its source and context clear.