Blog/Open source TTS models
Models · November 2026

Best open source text-to-speech models in November 2026: arena ranks, licenses, VRAM and speed

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

The best open-weight text-to-speech model on 7 November 2026 is Fish Audio Drama 3, seventh of 104 on the Artificial Analysis provider-voice leaderboard at 1275 Elo, 64 points behind ElevenLabs Eleven v4 (1339). Its license allows commercial use only below USD 500,000 annual revenue and never as a hosted service. The best model you can ship under MIT or Apache 2.0 is Ming-3.0-tts-0.5B (1187 Elo, effectively tied with Drama 3 on cloned voices); for a tiny CPU-friendly model, Kokoro-82M (1064) and Phonon 31M (1114) are Apache 2.0. Most of these need under 10 GB of memory; Drama 3 needs more than a 32 GB card.

ModelProvider voice: rank, EloCloned voice: rank, EloWeights licenseCommercial useSource
Eleven v4 (closed, for reference)1 of 104, 13394 of 49, 1153Closed APIPaid API onlyArtificial Analysis
Fish Audio Drama 37, 127515, 1073Fish Audio Community License 1.0Under USD 500k revenue, no hosted serviceLicense file
Breeze TTS 214, 122828, 1019BreezeBlue Research and Non-CommercialNoModel card
Ming-3.0-tts-0.5B25, 118717, 1067MITYesModel card
Phonon 31M (English)41, 1114not listedApache 2.0YesModel card
Fish Audio S2 Pro44, 110831, 1010Fish Audio Research LicenseSeparate licenseModel card
Voxtral 4B TTS 260351, 109232, 1007CC BY-NC 4.0NoModel card
Magpie TTS Multilingual 357M62, 1064not listedNVIDIA Open Model LicenseYesModel card
Kokoro-82M v1.063, 1064not listedApache 2.0YesModel card
Higgs Audio V3 TTS 4B78, 103940, 964Boson Higgs TTS 3 Research and Non-CommercialNo (creator grant)Model card
Chatterbox (English)85, 102343, 943MITYesModel card
XTTS v296, 91548, 849Coqui Public Model LicenseNoLicense file

Ranks and Elo from the Artificial Analysis provider-voice board (104 models, each provider's own voices) and controlled-voice board (49 models, the same 8 cloned voices), both read on 7 November 2026; the pages show no update date. "Not listed" means the model is absent from that board. Licenses checked on each Hugging Face card or license file the same day. Not legal advice.

Which open TTS model ranks highest in November 2026?

Fish Audio Drama 3 ranks highest among open-weight models on both Artificial Analysis boards, and the gap to the best closed model is now small. On the provider-voice leaderboard, where listeners pick the more natural of two clips in blind pairs, Drama 3 scores 1275 (95% interval 1257 to 1293, 1,618 samples). Eleven v4 leads at 1339. By my calculation with the standard Elo formula, 1 / (1 + 10^(−64/400)), listeners would prefer Eleven v4 in about 59% of pairs. Against Ming-3.0-tts at 1187, the figure rises to 71%.

The controlled-voice board removes the advantage of a provider's best studio voices by making every model clone the same eight voices. There Drama 3 scores 1073 and Ming-3.0-tts 1067: six points apart, inside Drama 3's interval of 1058 to 1088. Call it a tie. HeyGen Voice leads that board at 1189.

The older TTS Arena V2 on Hugging Face has few new open models entered. In its leaderboard on 7 November 2026, Fish Audio's OpenAudio S2 sits 14th at 1513, but it links to the paid service rather than to weights. The best entry with downloadable weights is Chatterbox at 1468 (29th of 41 ranked), with Kokoro v1.0 at 1465; the top non-preliminary entry, a stealth model called Aurora, has 1559. Drama 3, Ming and Qwen3-TTS are not on it.

Arenas measure preference, not intelligibility. Hugging Face's Open TTS Leaderboard (30 September 2026) scores word error rate, speaker similarity and speed instead; its write-up names Kokoro-82M, Supertonic 3 and Fish S2 Pro as the lowest-WER English models, and Higgs TTS 3 and VoxCPM2 as best when given a reference voice. Its result files require a login, so I cannot quote numbers.

Which open source TTS models can you use commercially?

Only about half of the open-weight models near the top of the arenas allow commercial use without a separate deal. Sorted by what the license files say:

Code and weights often have different licenses, so read the weights license, not the GitHub badge. License patterns across LLM families are in open-weight model licenses.

Which models speak German, French, Polish or Czech?

Western European languages are well covered; Czech is not, among commercially licensed models. From the cards:

XTTS v2 also covers Czech and Hungarian under a non-commercial license. Drama 3 claims 83 languages without listing them, so test yours before relying on it.

How much VRAM does a TTS model need, and how fast is it?

Most open TTS models need 1 to 16 GB and run faster than real time on one GPU; only Drama 3 is large. Cards disagree on "RTF": Mistral, Boson, Fish, VoxCPM and IndexTTS divide compute time by audio length (lower is faster), faster-qwen3-tts inverts it. I converted every figure to "times real time", seconds of audio per second of compute.

ModelParamsWeight filesMemory statedSpeed, x real timeHardware, engineSourceDate
Fish Audio Drama 335.9B (3B active)41.5 GBnot stated7.6 (1 stream), 80 total at 64H200, SGLang-Omni or vLLM-OmniCard2026-10-29
Voxtral 4B TTS 2603~4B8.0 GB16 GB GPU9.7 (1), 106 total at 32H200, vLLM 0.18.0Card2026-03-31
Higgs Audio V3 TTS 4B4.65B9.3 GBnot stated6.8 (1), 61 total at 16H100, bf16Card2026-09-04
Fish Audio S2 Pro4.56B11.0 GBnot stated5.1H200Card2026-03-11
Breeze TTS 23.47B7.6 GB7.7 GiB (12 GB GPU)3.1H100, fast pathCard2026-08-25
VoxCPM22.29B5.0 GB~8 GB3.3; 7.7 with Nano-vLLMRTX 4090Card2026-08-18
Qwen3-TTS 1.7B Base1.93B4.5 GB~6 GB (forum)4.22 RTX 4090; 1.87 DGX Sparkfaster-qwen3-tts 0.5.4, CUDA graphsPyPI2026-10-05
Qwen3-TTS 0.6B Base0.91B2.5 GBnot stated4.78 RTX 4090; 2.56 DGX Sparkfaster-qwen3-tts 0.5.4, CUDA graphsPyPI2026-10-05
IndexTTS-2.5not statednot checkednot stated4.8RTX 4090, bf16GitHub2026-08-10
F5-TTS v1 Basenot statednot checkednot stated6.8 PyTorch; 24.9 TensorRT-LLML20, 16 NFEGitHub2025-03-12
VibeVoice-Realtime-0.5B0.5B2.0 GBnot stated2.1DGX Spark, PyTorch cu130NVIDIA forum2026-01-04
Kokoro-82M v1.082M0.36 GBruns on CPU35 to 100RTX 4060 Ti 16 GB, Kokoro-FastAPIGitHubchecked 2026-11-07
Pocket TTS100Mper language2 CPU cores~6 on CPUMacBook Air M4Card2026-01-13

Parameters and weight-file sizes from the Hugging Face API for each repo on 7 November 2026; speeds are the publishers' own figures except the two forum rows, which are single-user reports. "Total at N" is aggregate speed with N concurrent requests, by my calculation N ÷ RTF from the card's table. The ~6 GB figure for Qwen3-TTS 1.7B is from martinB78 on the NVIDIA forum, 14 April 2026.

Two patterns. Single-stream speed on these autoregressive models depends on per-step overhead as much as on the GPU: in the faster-qwen3-tts table an H100 (3.30x for the 1.7B model) trails an RTX 4090 (4.22x), and CUDA graphs alone lift the 4090 from 0.82x. And batching multiplies throughput: Higgs V3 goes from 6.8x to 61x total at 16 concurrent requests on one H100. For a narration backlog, run requests in parallel, as for batch LLM jobs.

What does an hour of synthetic audio cost on your own GPU?

Under a dollar per finished audio hour on a rented GPU, even single-stream. The formula: compute minutes per audio hour = 60 ÷ (times real time), cost = hourly rate × minutes ÷ 60. At GPUwerk's rates (I run it: $0.79 per hour for a DGX Spark, $0.89 for the RTX 5090), by my calculation:

For scale, the Kokoro card puts API pricing for the same model under $0.06 per audio hour. A 10-hour audiobook through Qwen3-TTS 1.7B on one Spark stream takes about 5.3 hours, roughly $4.22; parallel requests should cut that, but no Spark concurrency figure for TTS is published.

Does TTS run on a DGX Spark or an RTX 5090?

Yes on both, with one exception: Drama 3's 41.5 GB of weights do not fit in the RTX 5090's 32 GB, and every other model here does. On the Spark's 128 GB, any of them fits next to a large LLM and a speech recognizer (budgeting shared memory is covered in GPU memory management for multiple models; the ASR side is in speech-to-text on a DGX Spark).

Per stream, the Spark runs Qwen3-TTS 1.7B at about 45% of an RTX 4090's speed (1.87x against 4.22x), still fast enough for a live voice agent. logosflux on the NVIDIA forum built a full pipeline on one Spark (whisper.cpp large-v3-turbo, Ollama with Llama 3.2 3B, VibeVoice-Realtime) and reports about 766 ms to first audio.

The ARM64 snag is torchaudio. In the Qwen3-TTS thread, pip pulled x86 or CPU-only wheels inside NVIDIA's PyTorch container; nav-intel fixed it on 10 February 2026 by reinstalling torch and torchaudio from the PyTorch cu130 index. For Kokoro, the Kokoro-FastAPI maintainer said on 27 October 2026 that the published GPU image is multi-arch with linux/arm64, so a local build is likely unnecessary. The general checklist is in five checks to run a new model on GB10.

Nobody has published a Drama 3 run on a Spark. From the card's numbers (927.6 frames per second across 16 streams at RTF 0.373), I calculate 21.6 frames per second of audio. If each frame reads the 3B active parameters at 1 byte (FP8) to 2 bytes (BF16), that is 3 to 6 GB per frame, and the Spark's 273 GB/s caps one stream at 45 to 91 frames per second: 2.1x to 4.2x real time. A bandwidth ceiling, not a prediction.

What do voice cloning and the EU AI Act require?

The model terms require the speaker's consent, and EU law adds marking and disclosure of synthetic audio. Publishers are tightening too: Microsoft removed the VibeVoice TTS code on 5 September 2025 after misuse, NVIDIA dropped zero-shot cloning from Magpie Multilingual "for security reasons", and Pocket TTS terms require explicit, lawful consent.

The AI Act adds two duties in Article 50. Paragraph 2: providers of systems generating synthetic audio must mark outputs in a machine-readable format so they are detectable as artificially generated. Paragraph 4: deployers who generate a deep fake must disclose it, with a lighter duty for evidently artistic or satirical work. Article 113 of the published text applies these from 2 August 2026. Chatterbox embeds a PerTh watermark in every file, the kind of marking paragraph 2 describes, though nothing I read says any given watermark satisfies it. For amendments since 2024 and what self-hosting changes, see the EU AI Act and open-weight models and GDPR and the AI Act on dedicated hardware. Not legal advice.

Which open TTS model should you pick?

For a commercial product in English, German, French or Spanish, I'd start with Ming-3.0-tts-0.5B: MIT license, the highest permissively licensed model on both Artificial Analysis boards, under 3 GB of weights. Its card gives no speed figures, so measure it first.

I would skip Chatterbox for new projects: Resemble marked the repo retired research in August 2026. To compare a Spark with an RTX 5090 for the rest of the stack, the trade-off is memory against per-stream speed, as for LLMs.

FAQ

What is the best open source text-to-speech model in November 2026?

On the Artificial Analysis provider-voice leaderboard read on 7 November 2026, Fish Audio Drama 3 is the top open-weight model at 1275 Elo, seventh of 104. Its community license stops at USD 500,000 annual revenue and excludes hosted services. The best MIT or Apache model there is Ming-3.0-tts-0.5B at 1187. ElevenLabs Eleven v4 leads the board at 1339.

Which open source TTS models can be used commercially?

From the licenses on their Hugging Face cards, checked 7 November 2026: Kokoro-82M, Qwen3-TTS, VoxCPM2, Phonon 31M, Dia2 and Sesame CSM (Apache 2.0), Ming-3.0-tts and Chatterbox (MIT), Kyutai Pocket TTS (CC-BY 4.0), and NVIDIA Magpie Multilingual (NVIDIA Open Model License). Voxtral TTS, Fish S2 Pro, Higgs Audio V3, Breeze TTS 2, F5-TTS, OmniVoice and XTTS v2 are non-commercial.

How much VRAM does a local TTS model need?

Most open TTS models are 0.1B to 5B parameters. Kokoro-82M runs on a CPU. Qwen3-TTS 1.7B used about 6 GB on a DGX Spark in one NVIDIA forum report, Breeze TTS 2 needs about 7.7 GiB, VoxCPM2 about 8 GB, and Voxtral 4B TTS a 16 GB GPU. Fish Audio Drama 3 has 41.5 GB of weight files, too much for a 32 GB RTX 5090.

Can a DGX Spark run text-to-speech in real time?

Yes, for one or two streams. The faster-qwen3-tts 0.5.4 benchmarks show Qwen3-TTS 1.7B at 1.87 times real time and the 0.6B model at 2.56 times on a DGX Spark with CUDA graphs, against 4.22 and 4.78 times on an RTX 4090. A forum user measured VibeVoice-Realtime-0.5B at about 2.1 times real time on a Spark.

Which open TTS models support Czech or Polish?

Polish is listed by Chatterbox Multilingual V3 (MIT), VoxCPM2 (Apache 2.0), Higgs Audio V3, XTTS v2 and OpenAudio S1-mini, the last three non-commercial. Czech is listed by Higgs Audio V3 (in its top tier, WER under 5) and XTTS v2, both non-commercial. No permissively licensed model I checked on 7 November 2026 lists Czech officially; Kyutai's Pocket TTS card links a community Czech model.

Do I need consent to clone a voice with an open TTS model?

Yes. Kyutai's Pocket TTS terms require explicit, lawful consent, and Sesame's CSM card forbids mimicking real people without it. In the EU, Article 50 of the AI Act (Regulation 2024/1689) also requires deployers to disclose deep fakes and providers to mark synthetic audio as machine-detectable, from 2 August 2026 in the published text. This is not legal advice.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

Related pages

First top-up: pay $20, get $30 in credit

Run speech recognition, an LLM and TTS on one machine.

Rent a 128 GB DGX Spark by the minute to test a voice pipeline, including models like Fish Audio Drama 3 that do not fit in 32 GB.

A dedicated DGX Spark in the EU, $0.79/hour, billed per minute. $30 of credit is about 37 hours of compute.

Create account Or rent the RTX 5090 for small TTS models