Best open-weight text-to-speech models
Best open-weight text-to-speech models is a comparison of text-to-speech (TTS) models whose trained weights can be downloaded and run on a user's own hardware, including models for voice cloning and multi-speaker dialogue. It covers models available as of November 2026. The models are grouped by the license attached to the weights, because that license decides whether a downloaded model may be used in a product, and by documented capability. The article does not rank them. Where it reports a position on a public leaderboard, it names the leaderboard, the date the figure was read and what the leaderboard measures.
On November 19, 2026, the Artificial Analysis Speech Arena leaderboard listed 20 models with downloadable weights among 105 TTS models. The highest placed, Fish Audio Drama 3, was seventh overall, and the three highest-placed open-weight entries were all flagged by Artificial Analysis as not licensed for commercial use.[1] On the same day the most downloaded repository tagged text-to-speech on Hugging Face was Kokoro-82M, an 82-million-parameter model released under the Apache License 2.0.[2][3]
Terminology and scope
This article uses "open-weight" for any model whose weights can be downloaded, whatever the license. The term describes availability of the weights only and does not imply that a model meets the Open Source AI Definition; see open weights and open-source AI. Some developers use the same word: the Kokoro model card calls Kokoro "an open-weight TTS model with 82 million parameters", and the Zonos repository calls Zonos-v0.1 "a leading open-weight text-to-speech model".[3][4] Others describe their releases as open source. Resemble AI's model card calls Chatterbox and Chatterbox Multilingual "production-grade open source TTS models", Canopy Labs calls Orpheus a "SOTA open-source text-to-speech system", and the Parler-TTS card calls that project a "fully open-source" release because its datasets, pre-processing code, training code and weights are all public.[5][6][7]
For each model the tables below give the license of the weights and the license of the inference code separately, because the two often differ. Hosted, closed speech products are covered in best AI voice generators, and license families in general are covered in open-weight license comparison and list of open-weight models. Companion comparisons cover image models and video models, and release dates across modalities are collected in the timeline of open-weight models.
Selection criteria
A model is included if, on November 19, 2026, all of the following held:
- its weights could be downloaded from a public repository, with or without a click-through access form;
- the license of the weights could be read in the model card, the repository metadata or a license file;
- there was independent evidence that the model is used or has been evaluated: a row on the Artificial Analysis Speech Arena leaderboard or on TTS Arena V2, or a download count returned by the Hugging Face Hub on that date.
Tables are in alphabetical order by model name. License tags, gating status and repository creation dates come from the Hugging Face Hub API as read on that date.[2] Dates are the developer's stated release date where one was found. A date marked (r) is the creation date of the model's Hugging Face repository as returned by the Hub API, which can differ from the public release: the Sesame CSM-1B repository was created on March 6, 2025 and Sesame's README dates the release to March 13, 2025, and the Voxtral TTS repository was created in November 2025 while its first public commit is dated March 26, 2026, three days after Mistral announced the model.[2][8][73]
Several repositories used in this article were gated on that date, meaning that Hugging Face requires a logged-in user to accept access conditions before downloading: sesame/csm-1b, canopylabs/orpheus-3b-0.1-ft, kyutai/pocket-tts, neuphonic/neutts-air, neuphonic/neutts-nano, neuphonic/neutts-2e, fishaudio/s1-mini and Gradium's six phonon-40m repositories.[2] Their model cards and license files could not be read without logging in, so the weights license given here for each of them is the repository's public metadata tag. A developer statement that could be read corroborates the tag only in two cases: the license table in Neuphonic's GitHub README, for the NeuTTS models, and the model table on the public phonon-31m-en card, for the 40M Phonon models.[21][23] For CSM-1B, Orpheus, Pocket TTS and S1-mini the tag is the only statement of the weights license that was read; the license files in the developers' GitHub repositories cover the code.[2][8][6][17]
One entry on the Artificial Analysis list, Step Audio EditX, is left out of the license tables because the second condition was not met: its Hugging Face repository carries no license tag, and the license section of its card names a license (Apache 2.0) for the code only.[2][67]
Independent evidence
Artificial Analysis Speech Arena
Artificial Analysis ranks TTS models with an Elo rating derived from blind votes: users hear two speech samples generated from the same text and choose the one that sounds more natural.[1] The site labels a model "open weights" if it can be self-hosted "using downloadable model weights or a provider-supplied inference runtime", and "open weights commercial" if its published license permits commercial use, including use subject to conditions, but not if a separately negotiated commercial agreement is required.[1] On November 19, 2026 the leaderboard had 105 models, led by Eleven v4 from ElevenLabs at 1,341, a score shared with Eleven v4 Turbo. Artificial Analysis publishes a 95% confidence interval and a rank range with each score, so the positions below follow the point estimates and neighboring entries can overlap: Drama 3 was shown at 1,279 plus or minus 16, with a rank range of 5 to 9. These 20 entries carried the open-weights label:[1]
| Rank (of 105) | Model | Elo | Commercial flag (Artificial Analysis) |
|---|---|---|---|
| 7 | Fish Audio Drama 3 | 1279 | No |
| 13 | Fish Audio Drama 3 Preview | 1237 | No |
| 14 | Breeze TTS 2 | 1229 | No |
| 25 | Ming-3.0-tts | 1182 | Yes |
| 44 | Phonon 31M (English) | 1110 | Yes |
| 47 | Step Audio EditX (Mar 2026) | 1101 | No |
| 48 | Fish Audio S2 Pro | 1101 | No |
| 50 | Voxtral TTS | 1094 | No |
| 63 | Magpie-Multilingual 357M (Feb 2026) | 1063 | Yes |
| 67 | Kokoro 82M v1.0 | 1060 | Yes |
| 73 | Maya1 | 1047 | Yes |
| 76 | OpenAudio S1 Mini | 1043 | No |
| 79 | Higgs Audio V3 TTS | 1035 | No |
| 86 | Chatterbox | 1020 | Yes |
| 88 | Zonos-v0.1 | 1000 | Yes |
| 90 | OpenVoice v2 | 955 | Yes |
| 91 | VibeVoice 1.5B | 954 | No |
| 97 | XTTS v2 | 915 | No |
| 101 | StyleTTS 2 | 888 | Yes |
| 104 | MetaVoice v1 | 845 | Yes |
Two rows refer to one repository: Artificial Analysis lists Fish Audio Drama 3 Preview and Fish Audio Drama 3 as separate entries, both linked to fishaudio/drama-3, whose weight files were replaced on October 29, 2026. "Higgs Audio V3 TTS" is the leaderboard's name for the model that Boson AI's card calls Higgs TTS 3.[1][2][41]
TTS Arena V2
TTS Arena V2 is a Hugging Face Space in which visitors compare two anonymous voices reading the same prompt and vote for the one that sounds more human. Its documentation says the board is a single Bradley-Terry fit over every recorded matchup, ordered by a conservative lower bound, that a model needs 100 votes to appear and is marked preliminary until it has 300, and that the main board uses English prompts.[9] On November 19, 2026 the board ranked 44 models. The board has its own open flag, and among entries with at least 300 votes it marked three: Kokoro v1.0, Veena and CosyVoice 2.0. The table below adds four entries that the board does not flag as open but whose names match model families with downloadable weights covered in this article. That matching is this article's, not the board's.[9]
| Rank (of 44) | Entry | Score | Votes | Flagged open by the board |
|---|---|---|---|---|
| 12 | OpenAudio S2 | 1556 | 401 | No |
| 25 | Chatterbox | 1490 | 2548 | No |
| 27 | Kokoro v1.0 | 1477 | 1850 | Yes |
| 32 | Magpie Multilingual | 1446 | 1622 | No |
| 35 | Maya 1 | 1412 | 586 | No |
| 37 | Veena | 1392 | 1212 | Yes |
| 40 | CosyVoice 2.0 | 1380 | 426 | Yes |
The four unflagged entries link to the developers' own sites or hosted services (Fish Audio for OpenAudio S2, Resemble AI for Chatterbox, NVIDIA's hosted catalog for Magpie Multilingual and Maya Research for Maya 1), not to weights repositories, so the system that was voted on may not be the downloadable checkpoint. StyleTTS 2 and Spark TTS were also listed, both marked preliminary with fewer than 300 votes.[9]
Positions on both boards move as votes arrive and as models are added. The Breeze TTS 2 model card, for example, says the model "ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard", while the board read on November 19, 2026 placed two Fish Audio Drama 3 entries above it.[10][1] Scores from the two boards cannot be compared with each other: they come from different voters, prompts, model sets and rating methods.[1][9]
Hugging Face downloads
Hugging Face counts a download server-side each time one of a repository's designated query files is requested.[11] The Hub's download figure, shown on model pages as "Downloads last month", counts file requests, not listener preferences, and it is kept per repository, so community conversions of a model have their own counts. Figures below were read from the Hub API on November 19, 2026 for a selection of the developers' own repositories. The table is not a complete ranking: community conversions and other checkpoints of the same families are left out.[2]
| Repository | Downloads | Weights license |
|---|---|---|
| hexgrad/Kokoro-82M | 13,092,988 | Apache-2.0 |
| coqui/XTTS-v2 | 5,095,820 | Coqui Public Model License |
| k2-fsa/OmniVoice | 4,210,382 | CC BY-NC |
| Qwen/Qwen3-TTS-12Hz-1.7B-Base | 4,061,594 | Apache-2.0 |
| ResembleAI/chatterbox | 2,577,323 | MIT |
| SWivid/F5-TTS | 1,339,981 | CC BY-NC 4.0 |
| openbmb/VoxCPM2 | 976,971 | Apache-2.0 |
| gradium/phonon-31m-en | 334,124 | Apache-2.0 |
| microsoft/VibeVoice-1.5B | 215,472 | MIT tag, research-only card |
| FunAudioLLM/Fun-CosyVoice3-0.5B-2512 | 157,538 | Apache-2.0 |
License tiers
Permissive licenses
The models in this tier have weights under the Apache License 2.0, the MIT License or CC BY 4.0 (see Creative Commons licenses), none of which restricts commercial use.
| Model | Developer | Weights license | Code license | Size | Languages (developer) | Date |
|---|---|---|---|---|---|---|
| Chatterbox (original, Multilingual V3, Turbo, Nano)[12][5] | Resemble AI | MIT | MIT | 500M; 350M (Turbo); 110M (Nano) | English; 23 (Multilingual) | 2025-04-24 (r) |
| CosyVoice 2 and Fun-CosyVoice 3[13] | Alibaba (FunAudioLLM) | Apache-2.0 | Apache-2.0 | 0.5B | 9, plus 18+ Chinese dialects (version 3) | 2024-12; 2025-12 |
| CSM-1B[8] | Sesame | Apache-2.0 (gated) | Apache-2.0 | 1B | English | 2025-03-13 |
| Dia 1.6B and Dia2[14][15] | Nari Labs | Apache-2.0 | Apache-2.0 | 1.6B; 1B and 2B (Dia2) | English | 2025-04-20 (r); 2025-11-15 (r) |
| Kokoro-82M[3] | hexgrad | Apache-2.0 | Apache-2.0 | 82M | 8 (v1.0) | 2025-01-27 (v1.0) |
| Kyutai TTS 1.6B and Pocket TTS[16][17] | Kyutai | CC BY 4.0 (Pocket TTS gated) | Apache-2.0; MIT | 1.8B; 100M | English, French; 7 | 2025-06-30 (r); 2026-01 (Pocket TTS report; repository created 2025-12-29) |
| Maya-1, Maya 2 Native and Veena[18][63][64] | Maya Research | Apache-2.0 | MIT notice on the streaming example (Maya-1); Apache-2.0 (Maya 2 Native) | 3.3B; 2.5B; 3B | No language list on card (Maya-1); 11 Indian languages and English (Maya 2 Native); Hindi and English (Veena) | 2025-10-18 (r); 2026-10-25; 2025-06-24 (r) |
| MetaVoice-1B[65] | MetaVoice | Apache-2.0 | Apache-2.0 | 1.2B | English | 2024-02-06 (r) |
| Ming-3.0-tts-0.5B[19] | inclusionAI (Ant Group) | MIT | MIT | 0.5B | 7, plus 2 Chinese dialects | 2026-09-22 (r); code repository created 2026-10-19 |
| MOSS-TTS family[20] | MOSI.AI and OpenMOSS team | Apache-2.0 | Apache-2.0 | About 100M (Nano) to 8B | 20; 31 (v1.5) | 2026-02-10 |
| NeuTTS Air[21] | Neuphonic | Apache-2.0 (gated) | Repository license file is the NeuTTS Open License | About 360M active | English | 2025-09-15 (r) |
| OpenVoice V2[22] | MyShell | MIT | MIT | Not stated | English, Spanish, French, Chinese, Japanese, Korean | 2024-04 |
| Orpheus TTS[6] | Canopy Labs | Apache-2.0 (gated) | Apache-2.0 | 3B | English; multilingual research preview | 2025-03-17 (r) |
| Parler-TTS Mini v1 and Large v1[7] | Hugging Face (Parler-TTS project) | Apache-2.0 | Apache-2.0 | 2.2B (Large) | English | 2024-06-26 (r); 2024-08-08 (r) |
| Phonon 31M English[23] | Gradium | Apache-2.0 | MIT or Apache-2.0 (runtime) | 31M | English | 2026-10-19 (r) |
| Qwen3-TTS[24] | Alibaba (Qwen team) | Apache-2.0 | Apache-2.0 | 0.6B; 1.7B | 10 | 2026-01-22 |
| VoxCPM2[25] | OpenBMB | Apache-2.0 | Apache-2.0 | 2B | 30 | 2026-04-03 (r) |
| Zonos-v0.1 and ZONOS3-Mini[4][26][68] | Zyphra | Apache-2.0 | Apache-2.0 (Zonos repository); MIT (ZONOS3-Mini repository) | 1.6B; 800M total, about 100M active | 5; 43 | 2025-02-06 (r); 2026-11-09 (r) |
Some entries in this tier carry additional terms or dependencies. CSM-1B uses a Llama backbone, and Sesame's instructions require access to the meta-llama/Llama-3.2-1B repository as well as its own; the Orpheus repository on Hugging Face lists meta-llama/Llama-3.2-3B-Instruct as its base model.[8][2] The Fun-CosyVoice repository carries an Apache-2.0 license and a disclaimer that its content is "for academic purposes only".[13] Resemble AI's README states that Chatterbox is no longer maintained (see below).[12] Only the 31M English Phonon model is Apache-2.0: Gradium's 40M English, French, German, Spanish, Portuguese and multilingual Phonon repositories are tagged CC BY-NC 4.0.[23][2] The Maya-1 card says the repository declares Apache 2.0 in its metadata while its streaming example "separately carries an MIT notice".[18] Two further repositories from developers in this table are also tagged Apache-2.0 on the Hub: hexgrad/Kokoro-82M-v1.1-zh (created February 27, 2025) and Zyphra/ZONOS2 (created June 11, 2026).[2]
Conditional and custom licenses
The licenses and cards in this tier attach conditions of their own: revenue or user thresholds, use restrictions that must be passed on, disclosure terms or, for VibeVoice, a research-only statement on the card beside an MIT tag.
| Model | Developer | Weights license | Main condition | Code license | Date |
|---|---|---|---|---|---|
| Fish Audio Drama 3[27][28] | Fish Audio | Fish Audio Community License 1.0 | Commercial use only while annual revenue is below USD 500,000; no license to offer a hosted service | Same license | 2026-10-13 |
| Higgs TTS 2 (formerly Higgs Audio v2)[29][30] | Boson AI | Boson Higgs Audio 2 Community License | Expanded license required above 100,000 annual active users | Apache-2.0 | 2025-07-01 (r) |
| IndexTTS-2 and IndexTTS-2.5[31][32] | bilibili IndexTeam | bilibili Model Use License | Separate license above 100 million monthly active users or RMB 1 billion annual revenue; no use to improve other AI models except non-commercial ones | Custom (no standard license detected by GitHub) | 2025-09-08; 2026-08-10 |
| Magpie TTS Multilingual 357M[33][70] | NVIDIA | NVIDIA Open Model License | Card: "This model is ready for commercial use." | Apache-2.0 (NeMo Speech) | 2026-01 (v2512); 2026-07-21 (v2607) |
| NeuTTS Nano and NeuTTS-2E[21][34] | Neuphonic | NeuTTS Open License 1.0 according to Neuphonic's README (both gated; Hub tags are "other" for Nano and Apache-2.0 for 2E) | Commercial use is licensed only while annual revenue is below USD 5 million; above that a paid license is required | Same license file | 2025-11-18 (r); 2026-07-21 (r) |
| StyleTTS 2[66] | Yinghao Aaron Li and co-authors | No license tag; README terms for the pre-trained models | Listeners must be told the speech is synthesized unless the user has permission to use the voice | MIT | 2023-11-21 (r) |
| Supertonic 3[35][62] | Supertone | BigScience OpenRAIL-M | Use restrictions in Attachment A, which must be passed on to downstream users | MIT (sample code; repository archived) | 2026-04-29 |
| VibeVoice-1.5B and VibeVoice-Realtime-0.5B[36][37][71] | Microsoft | MIT (tag) | Card limits the model to research use | MIT | 2025-08-25; 2025-12-03 |
For VibeVoice the license tag and the model card point in different directions. Both repositories are tagged MIT, but the VibeVoice-1.5B card says the model "is limited to research purpose use" and adds: "We do not recommend using VibeVoice in commercial or real-world applications without further testing and development."[36] The VibeVoice-Realtime-0.5B card likewise says the model "is limited to research purposes".[71] Artificial Analysis flags VibeVoice 1.5B as not licensed for commercial use.[1]
Neuphonic's documents and one Hub tag disagree. The README lists NeuTTS Air under Apache 2.0 and the Nano models and NeuTTS-2E under the NeuTTS Open License 1.0, while the gated neuphonic/neutts-2e repository is tagged Apache-2.0 on the Hub; this article follows the README. The license defines its threshold as "annual revenue of 5 million United States dollars ($5,000,000) or more".[21][34][2]
Supertone has ended work on Supertonic. The GitHub repository now sits in a supertone-oss-archive organization and opens with the notice "This repository is archived. Development and support have ended." Its README says the code and model weights keep their licenses. The Supertone/supertonic-3 repository on Hugging Face was still online on November 19, 2026, alongside an archive copy.[62][2]
Boson AI's card says that Higgs Audio V2 has been renamed Higgs TTS 2; the license file keeps the earlier name. StyleTTS 2's README puts the code under the MIT License and attaches a separate condition to the pre-trained models, which it says applies when the reference voice is not in the training set.[29][66]
The bilibili license names "bilibili indextts2" as the licensed model, and the copy shipped in the IndexTTS-2.5 repository uses the same wording. It also states that the Chinese-language version prevails over the English one.[32]
Non-commercial licenses
The weights in this tier are licensed for research or non-commercial use only, even where the code is under a permissive license.
| Model | Developer | Weights license | Code license | Notes | Date |
|---|---|---|---|---|---|
| Breeze TTS 2[10][38] | BreezeBlue | BreezeBlue Research and Non-Commercial License 1.1 | Apache-2.0 | English and Chinese; license also covers self-hosted outputs | 2026-08-25 |
| F5-TTS[39] | SWivid | CC BY-NC 4.0 | MIT | Restriction attributed to the Emilia training data | 2024-10-08 |
| Fish Audio S2 Pro[40] | Fish Audio | Fish Audio Research License | Fish Audio Community License | 80+ languages; Fish Speech 1.5 and S1-mini are CC BY-NC-SA 4.0 | 2026-03-09 (r) |
| Higgs TTS 3[41] | Boson AI | Boson Higgs TTS 3 Research and Non-Commercial License | Served with SGLang-Omni | 100+ languages; Creator Use Grant for published creative content | 2026-06-04 (r) |
| OmniVoice[42] | k2-fsa | CC BY-NC | Apache-2.0 | 600+ languages; restriction attributed to training data such as Emilia | 2026-03-30 (r) |
| Spark-TTS 0.5B[43][72] | SparkAudio | CC BY-NC-SA 4.0 | Apache-2.0 | Chinese and English | 2025-02-27 (r) |
| Voxtral TTS 4B[44][73] | Mistral AI | CC BY-NC 4.0 | Served with vLLM-Omni | 9 languages; license inherited from the reference voices | 2026-03-23 (announced); weights commit 2026-03-26 |
| XTTS-v2[45][46] | Coqui | Coqui Public Model License 1.0.0 | MPL-2.0 | 17 languages | 2023-10-31 (r) |
Developers give two kinds of reason for these restrictions. F5-TTS and OmniVoice both publish code under a permissive license and say the weights are CC BY-NC because of their training data, naming the Emilia dataset.[39][42] Mistral says the reference voices supplied with Voxtral TTS are CC BY-NC 4.0 and that this is "the license that the model inherits".[44] The Coqui Public Model License says that it "allows only non-commercial use of a machine learning model and its outputs", so it reaches generated audio as well as the weights.[46] The Hugging Face metadata for OmniVoice carries no license tag; the non-commercial term appears only in the text of the card.[2][42] The Spark-TTS card records that the model's license "has been updated from Apache 2.0 to CC BY-NC-SA due to the licensing terms of some training data", while the code repository remains under Apache-2.0.[72][43]
Models by use case
The groupings below follow capabilities documented by each developer. They are not rankings, and inclusion in a group says nothing about output quality.
Small models and CPU inference
- Kokoro-82M has 82 million parameters and ships 54 preset voices across 8 languages in version 1.0.[3]
- Pocket TTS is described by Kyutai as designed to run on CPUs, with a 100M-parameter model and English models down to 1.9M parameters.[17]
- Phonon 31M has 31 million parameters excluding its audio tokenizer, and Gradium says it is small enough to run in a browser tab, on a phone or on a laptop CPU. It builds on Pocket TTS and Kyutai's Mimi codec.[23]
- Chatterbox-Nano is a 110M-parameter English model that Resemble AI says targets on-device and CPU inference.[12]
- MOSS-TTS-Nano is an approximately 100M-parameter model that its developers say supports streaming output on 4 CPU cores.[20]
- Supertonic 3 is distributed as ONNX assets of about 99M parameters, and Supertone says it does not require a GPU.[35]
- ZONOS3-Mini has 800M parameters in total with about 100M active per token, and Zyphra says it runs in real time on phones and laptops, including in a browser.[26]
Voice cloning
Many of the models above document zero-shot cloning, in which the model imitates a speaker from a reference recording without being fine-tuned on that voice.[6][43] Developers document different reference lengths:
| Model | Documented reference audio |
|---|---|
| Dia 1.6B | 5 to 10 seconds recommended[14] |
| Fish Audio Drama 3 | Card's example uses a 3.9-second clip with its transcript[27] |
| MetaVoice-1B | 30 seconds, for American and British voices[65] |
| NeuTTS Air and Nano | "as little as 3 seconds of audio"[21] |
| OmniVoice | 3 to 10 seconds[42] |
| Qwen3-TTS (Base) | "3-second rapid voice clone"[24] |
| XTTS-v2 | 6-second clip[45] |
| Zonos-v0.1 | README: 10 to 30 second speaker sample; Zyphra's release post: clips of between 5 and 30 seconds[4][69] |
| ZONOS3-Mini | Reference clip of up to 60 seconds, or a speaker embedding[26] |
Chatterbox, CosyVoice, Higgs TTS 3, IndexTTS-2.5, Maya 2 Native, Ming-3.0-tts, Orpheus, Pocket TTS, Spark-TTS and VoxCPM2 also document cloning from reference audio; VoxCPM2 and Breeze TTS 2 can additionally take the transcript of the reference.[12][13][41][31][63][19][6][17][43][25][10]
Several models do not clone from arbitrary audio. Kokoro ships preset voices. NVIDIA's Magpie card says the model uses five speaker voices and "is not intended for zero-shot voice cloning".[3][33] Maya-1 and Parler-TTS build a voice from a written description instead of a recording.[18][7] Kyutai TTS 1.6B accepts only pre-computed voice embeddings, a restriction Kyutai chose in place of watermarking.[16] The open Phonon model comes with a catalog of about 100 voices, and Gradium says instant voice cloning and voice design are available through its studio.[23] NeuTTS-2E is a fixed-speaker model with four voices, and the open-weight Supertonic 3 package includes fixed preset voice styles.[21][35] Some models, including Qwen3-TTS, VoxCPM2, Ming-3.0-tts and Breeze TTS 2, add "voice design", in which a new voice is created from a text description.[24][25][19][10]
Multi-speaker dialogue and long-form audio
- Dia generates dialogue from a transcript marked with [S1] and [S2] speaker tags and can produce nonverbal sounds; Dia2 is a streaming dialogue model limited to 2 minutes of English per generation.[14][15]
- VibeVoice-1.5B is documented as synthesizing up to 90 minutes of speech with up to 4 speakers, in English and Chinese.[36]
- MOSS-TTSD is the MOSS family's model for multi-speaker dialogue, podcasts and dubbing.[20]
- Fish Audio Drama 3 uses speaker tags numbered 0 to 15, so one text can carry up to 16 characters.[27]
- Sesame CSM takes earlier turns of a conversation, as text and audio for each speaker, as context for the next utterance.[8]
- Kyutai TTS 1.6B is documented as performing streaming generation "including dialogs".[16]
Multilingual coverage
Language counts are the developers' own and do not indicate equal quality across languages. Among models whose developers state nine or more languages, in descending order of the stated count: OmniVoice (600+), Higgs TTS 3 (100+), Fish Audio S2 Pro (80+), ZONOS3-Mini (43), MOSS-TTS v1.5 and Supertonic 3 (31 each), VoxCPM2 (30), Chatterbox Multilingual (23), XTTS-v2 (17), Magpie TTS Multilingual and Maya 2 Native (12 each), Qwen3-TTS (10), and Fun-CosyVoice 3 and Voxtral TTS (9 each).[42][41][40][26][20][35][25][12][45][33][63][24][13][44] Of the three models with the widest stated coverage, none has weights licensed for general commercial use. Sesame warns that CSM-1B has some capacity for non-English languages through data contamination but "likely won't do well".[8]
Real-time streaming
Latency and speed figures below are developers' statements with the hardware each one names. They were measured under different conditions and are not comparable with one another. Speed is also expressed in two ways: OpenBMB, Fish Audio and BreezeBlue give a real-time factor, the processing time relative to the duration of the audio, where lower is faster, while Zyphra, Nari Labs, Kyutai and Resemble AI give a multiple of real time, where higher is faster.
| Model | Developer statement | Hardware named |
|---|---|---|
| Breeze TTS 2 | Under 40 ms time to first audio; real-time factor 0.32 (warmed-up fast path)[10] | NVIDIA H100 |
| Chatterbox-Nano | 3x faster than real time[12] | 8 CPU cores |
| Dia 1.6B | 2.1x real time with bfloat16 and compilation, about 4.4 GB of VRAM[14] | RTX 4090 |
| Fish Audio Drama 3 | Median time to first audio 88.9 ms, mean real-time factor 0.131 at concurrency 1[27] | One NVIDIA H200 |
| Fish Audio S2 Pro | Real-time factor 0.195[40] | One NVIDIA H200 |
| Fun-CosyVoice 3 | Latency "as low as 150ms" in bi-streaming mode[13] | None named |
| Orpheus TTS | About 200 ms streaming latency, about 100 ms with input streaming[6] | None named |
| Pocket TTS | About 200 ms to the first audio chunk; about 6x real time on 2 CPU cores[17] | MacBook Air M4 CPU |
| Qwen3-TTS | End-to-end synthesis latency "as low as 97ms"[24] | None named |
| VoxCPM2 | Real-time factor about 0.3, or about 0.13 with Nano-vLLM[25] | RTX 4090 |
| Zonos-v0.1 | Real-time factor of about 2x[4] | RTX 4090 |
| ZONOS3-Mini | Median time to first audio 80 ms with 8 concurrent requests[26] | One H100 80GB |
Microsoft describes VibeVoice-Realtime-0.5B as a real-time model that accepts streaming text input, and the MOSS family includes a separate MOSS-TTS-Realtime model.[37][20]
License changes, withdrawals and company events
The events below changed the license, availability or maintenance status of models covered in this article.
| Date | Developer | Event |
|---|---|---|
| 2024-01-03 | Coqui | A discussion on the coqui-ai/TTS repository reported that Coqui's website announced the company was shutting down.[47] |
| 2024-01-05 | Coqui | Asked to relicense XTTS, a Coqui maintainer replied "Currently we don't have plans to change the license", and answered "most probably no" when asked whether new customers could still buy a commercial license.[48] |
| 2024-04 | MyShell | OpenVoice V1 and V2 placed under the MIT License.[22] |
| 2025-03-07 | SparkAudio | License of the Spark-TTS-0.5B weights changed from Apache 2.0 to CC BY-NC-SA 4.0 (date of the "Update License" commits on the Hub).[72][2] |
| 2025-09-05 | Microsoft | VibeVoice-TTS code removed from the GitHub repository after Microsoft "discovered instances where the tool was used in ways inconsistent with the stated intent".[37] |
| 2026-07-08 | Boson AI | Higgs TTS 3 license text updated; it adds a Creator Use Grant for monetized creative content while keeping other commercial use under a separate license.[41][2] |
| 2026-07-23 | Supertone | Notice that the Supertonic repository would be archived, with "no further development or official support for the open-source Supertonic models". The archived repository was last pushed on September 9, 2026.[62] |
| 2026-08-13 | Resemble AI | Chief executive's post says the company is "not selling voice AI to new customers".[52] |
| 2026-10-13 | Fish Audio | Fish Audio Community License 1.0 issued for Drama 3 and the fish-speech codebase; S2 Pro stays under the Fish Audio Research License.[28][49] |
| 2026-11-04 | Resemble AI | Notice added to the Chatterbox README: "As of August 2026, Chatterbox is retired research and is no longer maintained."[12][74] |
Coqui's original repository published its last release in December 2023.[50] The idiap/coqui-ai-TTS fork describes itself as a "Fork of the original, unmaintained repository" and published version 0.28.0 in October 2026. The code is under the Mozilla Public License 2.0, but the XTTS-v2 weights remain under the non-commercial Coqui Public Model License, and XTTS-v2 was still the second most downloaded developer repository in the table above.[51][46][2]
The Chatterbox README opens with a notice, added in a commit dated November 4, 2026: "As of August 2026, Chatterbox is retired research and is no longer maintained." It says Resemble AI does not sell text-to-speech or voice cloning products and that the code "stays public as an archive under the MIT license, with no support, updates, or paid tier".[12][74] Further down, the same README still carries an older paragraph pointing readers to a Resemble AI TTS service.[12] In a post dated August 13, 2026, chief executive Zohaib Ahmed wrote that the company was "not selling voice AI to new customers" and had put its effort into deepfake detection, and that its voice models "are open and free to use, but we won't be producing new ones for commercial sale". Of voice generation as a paid product, the post says: "We're supporting our existing customers but not taking new ones."[52]
The VibeVoice-1.5B card lists a larger VibeVoice-Large checkpoint as "Disabled". Microsoft released VibeVoice-Realtime-0.5B in the same repository on December 3, 2025.[36][37]
Repository metadata can lag behind such changes. On November 19, 2026 the Hugging Face tag on fishaudio/drama-3 still read "fish-audio-research-license", while the card and the license file in the repository both named the Fish Audio Community License.[2][27] The CosyVoice code repository, published under the FunAudioLLM name, now resolves on GitHub to QwenAudio/CosyVoice, and the Supertonic repository to supertone-oss-archive/supertonic.[13][62] Development of the Piper TTS engine moved from the MIT-licensed rhasspy/piper repository, now archived, to OHF-Voice/piper1-gpl, which is under the GNU GPL 3.0, a copyleft license.[53][54]
Misuse, consent and safeguards
The U.S. Federal Communications Commission said in February 2024 that voice cloning technology was being used in common robocall scams and could imitate the voices of celebrities, political candidates and close family members.[55] Synthetic audio that resembles a real person and would falsely appear authentic falls within the EU AI Act's definition of a deepfake.[56] Developers of open-weight models address misuse in three ways: terms of use, technical marking of output, and limits on what the released model can do.
License and model-card terms
Several custom licenses make consent a condition of use. The Fish Audio Community License forbids generating speech "that imitates the voice of an identifiable natural person without that person's prior consent", with separate provision for minors and deceased persons and an exception for a user who has "another legal right to do so".[28] Boson's Higgs TTS 3 license forbids imitating a real person without "explicit, verifiable consent", and BreezeBlue's license contains a similar clause.[41][38] The OpenRAIL-M license used for Supertonic 3 lists impersonating others without their consent among its use restrictions.[35]
For models under Apache-2.0 or MIT, the corresponding requests appear in the model card or README instead of the license text. Sesame's README lists, among uses it says it explicitly prohibits, generating "speech that mimics real individuals without their explicit consent"; Nari Labs lists producing audio resembling real individuals without permission among uses that are "strictly forbidden"; the VoxCPM README forbids use for impersonation, fraud or disinformation and recommends marking generated content; the Pocket TTS README prohibits "voice impersonation or cloning without explicit and lawful consent"; the Ming-3.0-tts card tells users not to impersonate a real person without consent; and the Maya 2 Native card tells users to obtain the speaker's permission before cloning a voice.[8][14][25][17][19][63] The Chatterbox README's disclaimer says: "Don't use this model to do bad things."[12]
Watermarking and disclosure
- Chatterbox applies Resemble AI's PerTh watermark to every generated file. Resemble describes the marks as imperceptible and as surviving MP3 compression and audio editing.[12]
- NeuTTS output includes a PerTh watermark by default. Neuphonic's README adds that watermarking is disabled, with a warning, when the package is installed from the repository with uv sync, because the Perth dependency is then missing.[21]
- For VibeVoice, Microsoft lists three mitigations: an audible disclaimer embedded in every synthesized file, an imperceptible watermark, and hashed logging of inference requests.[36]
- The Orpheus README suggests watermarking generated audio with Silent Cipher and links an implementation.[6]
- Kyutai states that its TTS model performs no watermarking, for two reasons: watermarking can easily be deactivated in a model whose code is public, and its early experiments found that the watermarks used by existing TTS systems were removed by encoding and decoding the audio with its Mimi codec.[16]
Law and regulation
Some cards place legal compliance on the user. Mistral's Voxtral TTS card tells users "You are responsible for complying with applicable laws and avoiding misuse."[44]
In the United States, the Federal Communications Commission announced on February 8, 2024 a unanimous Declaratory Ruling that calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act. The Commission said the ruling took effect immediately and gave state attorneys general new means to act against voice-cloning robocall scams.[55] The ruling itself (FCC 24-17, CG Docket No. 23-362, adopted February 2 and released February 8, 2024) confirms that the Act's restrictions on the use of an "artificial or prerecorded voice" encompass current AI technologies that generate human voices, so that such calls require the prior express consent of the called party unless an emergency purpose or an exemption applies.[75]
In the European Union, Article 50 of the AI Act (Regulation (EU) 2024/1689) requires providers of AI systems that generate synthetic audio, image, video or text to ensure the output is marked in a machine-readable format and detectable as artificially generated. It also requires deployers of a system that produces a deep fake to disclose that the content was artificially generated or manipulated. The regulation's definition of "deep fake" includes audio. Under Article 113, as published in the Official Journal on July 12, 2024, the regulation applies from August 2, 2026 apart from listed exceptions, and Article 50 is not among them.[56] Regulation (EU) 2026/1744 (the Digital Omnibus on AI), published in the Official Journal on July 24, 2026, amended the exceptions in Article 113 without adding Article 50 to them. It added Article 111(4), under which providers of systems generating synthetic audio, image, video or text that were placed on the market before August 2, 2026 "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026", and it replaced Article 50(7), on codes of practice.[76]
Running the models
Some models have lighter runtimes in addition to their reference code: the MOSS-TTS developers publish quantized GGUF weights and a llama.cpp pipeline with an ONNX audio codec, and say an 8B model then fits on an 8 GB GPU; Neuphonic publishes GGUF builds of NeuTTS.[20][21]
LocalAI exposes TTS through an API that its documentation describes as compatible with the OpenAI and ElevenLabs speech APIs. Its documentation lists dedicated backends for reference-audio cloning that include Chatterbox, Qwen3-TTS, Fish Speech, MOSS-TTS, NeuTTS, OmniVoice, Pocket TTS and VoxCPM, along with Coqui (for XTTS) and a VibeVoice backend.[57] General background on self-hosting is in local AI and self-hosting LLMs.
Open-weight speech recognition
Speech recognition, the reverse task, has its own open-weight models. Whisper from OpenAI is distributed from a GitHub repository under the MIT License, while the openai/whisper-large-v3 repository on Hugging Face is tagged Apache-2.0.[58][59] NVIDIA's Parakeet (parakeet-tdt-0.6b-v3) and Canary (canary-1b-v2) are both tagged CC BY 4.0.[60][61] See automatic speech recognition models and speech recognition.
Continue reading
- Voice AIVoice AI is the umbrella term for artificial intelligence systems that listen to, understand, and generate human speech.Mentioned twice above
- Text-to-Speech ModelsText-to-speech (TTS) models are machine learning systems that convert written text into spoken audio.Listed in See also
- Best Open-Source LLMsAs of November 17, 2026, the highest-scoring open-weight large language model is Moonshot AI's Kimi K3.1 Pro, whose weights Moonshot released on November 4.Listed in See also
- Best Local and On-Device LLMsAs of November 2026, the best local LLM to run offline depends almost entirely on how much memory you can give it, so this guide ranks picks by hardware tier.Listed in See also
- Open-source licenseAn open-source license is a software license that complies with the Open Source Definition (OSD), a ten-point standard maintained by the Open Source Initiative (OSI).Listed in See also
- Open coreOpen core is a business model in which a company publishes the main part of a software product under an open-source license and sells additional features under a proprietary licenseListed in See also
More from See also
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Artificial Analysis. "Text to Speech Leaderboard - Top AI Speech Models". Artificial Analysis, read November 19, 2026. artificialanalysis.ai/...leaderboard
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Hugging Face. "Models" (Hub listing and API metadata for text-to-speech repositories: downloads, license tags, gating, creation dates and commits). Hugging Face Hub, read November 19, 2026. huggingface.co/models
- ^1 ^2 ^3 ^4 ^5hexgrad. "Kokoro-82M" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Kokoro-82M
- ^1 ^2 ^3 ^4Zyphra. "Zonos" (README). GitHub. Accessed November 19, 2026. github.com/...Zonos
- ^1 ^2Resemble AI. "Chatterbox" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...chatterbox
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Canopy Labs. "Orpheus TTS" (README). GitHub. Accessed November 19, 2026. github.com/...Orpheus-TTS
- ^1 ^2 ^3Parler-TTS. "Parler-TTS Large v1" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...parler-tts-large-v1
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Sesame. "CSM" (README). GitHub. Accessed November 19, 2026. github.com/...csm
- ^1 ^2 ^3 ^4mrfakename; Srivastav, Vaibhav; Fourrier, Clémentine; Pouget, Lucain; Lacombe, Yoach; main; Gandhi, Sanchit; Passos, Apolinário; Cuenca, Pedro. "TTS Arena 2.0: Benchmarking Text-to-Speech Models in the Wild". Hugging Face, 2025; leaderboard and methodology read November 19, 2026. huggingface.co/...TTS-Arena-V2
- ^1 ^2 ^3 ^4 ^5BreezeBlue. "Breeze TTS 2" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Breeze-TTS-2
- ^Hugging Face. "Models Download Stats". Hugging Face Hub documentation. Accessed November 19, 2026. huggingface.co/...models-download-stats
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Resemble AI. "Chatterbox" (README). GitHub. Accessed November 19, 2026. github.com/...chatterbox
- ^1 ^2 ^3 ^4 ^5 ^6FunAudioLLM. "CosyVoice" (README; repository now at QwenAudio/CosyVoice). GitHub. Accessed November 19, 2026. github.com/...CosyVoice
- ^1 ^2 ^3 ^4 ^5Nari Labs. "Dia" (README). GitHub. Accessed November 19, 2026. github.com/...dia
- ^1 ^2Nari Labs. "Dia2" (README). GitHub. Accessed November 19, 2026. github.com/...dia2
- ^1 ^2 ^3 ^4Kyutai. "Model Card for Kyutai TTS" (kyutai/tts-1.6b-en_fr). Hugging Face. Accessed November 19, 2026. huggingface.co/...tts-1.6b-en_fr
- ^1 ^2 ^3 ^4 ^5 ^6Kyutai. "Pocket TTS" (README). GitHub. Accessed November 19, 2026. github.com/...pocket-tts
- ^1 ^2 ^3Maya Research. "Maya-1" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...maya1
- ^1 ^2 ^3 ^4inclusionAI. "Ming-3.0-tts-0.5B" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Ming-3.0-tts-0.5B
- ^1 ^2 ^3 ^4 ^5 ^6MOSI.AI and OpenMOSS team. "MOSS-TTS Family" (README). GitHub. Accessed November 19, 2026. github.com/...MOSS-TTS
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Neuphonic. "NeuTTS" (README). GitHub. Accessed November 19, 2026. github.com/...neutts
- ^1 ^2MyShell. "OpenVoice" (README). GitHub. Accessed November 19, 2026. github.com/...OpenVoice
- ^1 ^2 ^3 ^4 ^5Gradium. "phonon-31m English" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...phonon-31m-en
- ^1 ^2 ^3 ^4 ^5Qwen team, Alibaba Cloud. "Qwen3-TTS" (README). GitHub. Accessed November 19, 2026. github.com/...Qwen3-TTS
- ^1 ^2 ^3 ^4 ^5 ^6OpenBMB. "VoxCPM" (README). GitHub. Accessed November 19, 2026. github.com/...VoxCPM
- ^1 ^2 ^3 ^4 ^5Zyphra. "ZONOS3-Mini" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...ZONOS3-Mini
- ^1 ^2 ^3 ^4 ^5Fish Audio. "Fish Audio Drama 3" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...drama-3
- ^1 ^2 ^3Fish Audio. "Fish Audio Community License Agreement, Version 1.0". Hugging Face, last updated October 13, 2026. huggingface.co/...LICENSE
- ^1 ^2Boson AI. "Higgs TTS 2" (model card and Boson Higgs Audio 2 Community License Agreement, bosonai/higgs-tts-2-3b-base). Hugging Face. Accessed November 19, 2026. huggingface.co/...higgs-tts-2-3b-base
- ^Boson AI. "higgs-audio" (repository). GitHub. Accessed November 19, 2026. github.com/...higgs-audio
- ^1 ^2bilibili IndexTeam. "IndexTTS" (README). GitHub. Accessed November 19, 2026. github.com/...index-tts
- ^1 ^2bilibili. "bilibili Model Use License Agreement" (IndexTeam/IndexTTS-2.5). Hugging Face. Accessed November 19, 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3NVIDIA. "MagpieTTS Multilingual 357M" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...magpie_tts_multilingual_357m
- ^1 ^2Neuphonic. "NeuTTS Open License v1.0". GitHub, last updated December 11, 2025. github.com/...LICENSE
- ^1 ^2 ^3 ^4 ^5Supertone. "Supertonic 3" (model card and BigScience Open RAIL-M License). Hugging Face. Accessed November 19, 2026. huggingface.co/...supertonic-3
- ^1 ^2 ^3 ^4 ^5Microsoft. "VibeVoice-1.5B" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...VibeVoice-1.5B
- ^1 ^2 ^3 ^4Microsoft. "VibeVoice" (README). GitHub. Accessed November 19, 2026. github.com/...VibeVoice
- ^1 ^2BreezeBlue. "BreezeBlue Research and Non-Commercial License Agreement, Version 1.1". Hugging Face, last updated September 1, 2026. huggingface.co/...LICENSE
- ^1 ^2SWivid. "F5-TTS" (README). GitHub. Accessed November 19, 2026. github.com/...F5-TTS
- ^1 ^2 ^3Fish Audio. "Fish Audio S2 Pro" (model card and Fish Audio Research License, last updated March 7, 2026). Hugging Face. Accessed November 19, 2026. huggingface.co/...s2-pro
- ^1 ^2 ^3 ^4 ^5 ^6Boson AI. "Higgs TTS 3" (model card and Boson Higgs TTS 3 Research and Non-Commercial License, last updated July 8, 2026). Hugging Face. Accessed November 19, 2026. huggingface.co/...higgs-tts-3-4b
- ^1 ^2 ^3 ^4 ^5k2-fsa. "OmniVoice" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...OmniVoice
- ^1 ^2 ^3 ^4SparkAudio. "Spark-TTS" (README). GitHub. Accessed November 19, 2026. github.com/...Spark-TTS
- ^1 ^2 ^3 ^4Mistral AI. "Voxtral 4B TTS 2603" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Voxtral-4B-TTS-2603
- ^1 ^2 ^3Coqui. "XTTS-v2" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...XTTS-v2
- ^1 ^2 ^3Coqui. "Coqui Public Model License 1.0.0". Hugging Face (coqui/XTTS-v2). Accessed November 19, 2026. huggingface.co/...LICENSE.txt
- ^coqui-ai/TTS. "Coqui is shutting down." (discussion #3489). GitHub, January 3, 2024. github.com/...3489
- ^coqui-ai/TTS. "XTTS License After Shutdown" (issue #3490). GitHub, January 4, 2024. github.com/...3490
- ^Fish Audio. "Fish Speech" (README). GitHub. Accessed November 19, 2026. github.com/...fish-speech
- ^coqui-ai. "TTS" (repository and releases). GitHub. Accessed November 19, 2026. github.com/...TTS
- ^idiap/coqui-ai-TTS. "Coqui TTS" (README and releases). GitHub. Accessed November 19, 2026. github.com/...coqui-ai-TTS
- ^1 ^2Ahmed, Zohaib. "Why a former voice AI company went all in on deepfake detection". Resemble AI, August 13, 2026. resemble.ai/...y-went-all-in-on-deepfake-detection
- ^rhasspy. "Piper" (archived repository). GitHub. Accessed November 19, 2026. github.com/...piper
- ^OHF-Voice. "piper1-gpl" (repository). GitHub. Accessed November 19, 2026. github.com/...piper1-gpl
- ^1 ^2Federal Communications Commission. "FCC Makes AI-Generated Voices in Robocalls Illegal". News release, February 8, 2024. docs.fcc.gov/...DOC-400393A1.txt
- ^1 ^2European Parliament and Council of the European Union. "Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence". Official Journal of the European Union, L series, July 12, 2024. eur-lex.europa.eu/...HTML
- ^LocalAI. "Text to Audio (TTS)". LocalAI documentation (source file), GitHub. Accessed November 19, 2026. github.com/...text-to-audio.md
- ^OpenAI. "Whisper" (repository). GitHub. Accessed November 19, 2026. github.com/...whisper
- ^OpenAI. "whisper-large-v3" (model repository). Hugging Face. Accessed November 19, 2026. huggingface.co/...whisper-large-v3
- ^NVIDIA. "parakeet-tdt-0.6b-v3" (model repository). Hugging Face. Accessed November 19, 2026. huggingface.co/...parakeet-tdt-0.6b-v3
- ^NVIDIA. "canary-1b-v2" (model repository). Hugging Face. Accessed November 19, 2026. huggingface.co/...canary-1b-v2
- ^1 ^2 ^3 ^4Supertone. "Supertonic" (README of the archived repository). GitHub. Accessed November 19, 2026. github.com/...supertonic
- ^1 ^2 ^3 ^4Maya Research. "Maya 2 Native: Streaming Text-to-Speech for Indian Languages" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...maya2-native
- ^Maya Research. "Veena" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Veena
- ^1 ^2MetaVoice. "MetaVoice-1B" (README). GitHub. Accessed November 19, 2026. github.com/...metavoice-src
- ^1 ^2Li, Yinghao Aaron; Han, Cong; Raghavan, Vinay S.; Mischler, Gavin; Mesgarani, Nima. "StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models" (README). GitHub. Accessed November 19, 2026. github.com/...StyleTTS2
- ^StepFun. "Step-Audio-EditX" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Step-Audio-EditX
- ^Zyphra. "ZONOS3-Mini" (repository and license file). GitHub. Accessed November 19, 2026. github.com/...ZONOS3-Mini
- ^Zyphra. "Beta Release of Zonos-v0.1". Zyphra, February 10, 2025. zyphra.com/...beta-release-of-zonos-v0-1
- ^NVIDIA. "NeMo Speech" (repository and license file). GitHub. Accessed November 19, 2026. github.com/...Speech
- ^1 ^2Microsoft. "VibeVoice-Realtime-0.5B" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...VibeVoice-Realtime-0.5B
- ^1 ^2 ^3SparkAudio. "Spark-TTS-0.5B" (model card). Hugging Face. Accessed November 19, 2026. huggingface.co/...Spark-TTS-0.5B
- ^1 ^2Mistral AI. "Speaking of Voxtral". Mistral AI, March 23, 2026. mistral.ai/...voxtral-tts
- ^1 ^2Resemble AI. "Update README (#581)" (commit ab55923). GitHub, November 4, 2026. github.com/...923c27757c90cfcc407baa57dc25a7d25217
- ^Federal Communications Commission. "Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts: Declaratory Ruling". FCC 24-17, CG Docket No. 23-362, adopted February 2, 2024, released February 8, 2024. docs.fcc.gov/...FCC-24-17A1.pdf
- ^European Parliament and Council of the European Union. "Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI)". Official Journal of the European Union, L series, July 24, 2026. eur-lex.europa.eu/...HTML
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 7,350 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 19, 2026: 430 claims checked against 190 cited and independent sources; 7 material and 19 minor corrections applied before publication.
Cite this page: AI Wiki. "Best open-weight text-to-speech models." aiwiki.ai, updated 19 Nov 2026, fact-checked 19 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/best_open_weight_tts_models