Free audio AI models

Two rankings of open models you can use free: speech to text (they write down what you say) and text to speech (they read text aloud). Arena does not rank audio, so these come from two other public leaderboards, credited below.

Best speech to textQwen3 ASR 1.7B4.31% of words wrong · 2.04 B

Best text to speechkokoroscore 1211 · 34K votes

Open models ranked9049 speech to text, 41 text to speech

Speech to text (audio in)

Qwen3 ASR 1.7B 95.69 Hojo ASR V1 95.67 higgs audio v3 stt 95.61 granite speech 5.0 2b 95.58 canary qwen 2.5b 95.57 ARK ASR 3B 95.53 thewhisper large v3 turbo 95.46 ARK ASR 0.6B 95.44 granite speech 4.1 2b 95.38 MOSS Transcribe Diarize 95.36 cohere transcribe 03 2026 95.33 parakeet tdt 0.6b v2 (fast gpu asr) 95.33

909294

Words transcribed right (%), English, axis starts at 90. Higher is better. Average over 8 English test sets that every open model in the list has.

What this shows · How to read it

What it shows. The 12 open speech-to-text models that make the fewest mistakes on English recordings: meetings, phone calls, read books, accented speech.

How to read it. Longer bar, more words right. Differences under half a point are small; pick by size and speed too, in the table below.

#	Model	Words wrong	Size	Languages	Speed	License
1	Qwen3 ASR 1.7B Qwen	4.31%	2.04 B	52	836×	apache-2.0
2	Hojo ASR V1 HojoAI	4.33%	5.18 B	2	73×	apache-2.0
3	higgs audio v3 stt bosonai	4.39%	2.68 B	1	110×	apache-2.0
4	granite speech 5.0 2b ibm-granite	4.42%	2.31 B	6	545×	apache-2.0
5	canary qwen 2.5b nvidia	4.43%	2.5 B	—	882×	cc-by-4.0
6	ARK ASR 3B Edge0	4.47%	3.75 B	19	480×	apache-2.0
7	thewhisper large v3 turbo TheStageAI	4.54%	0.8 B	28	1K×	cc-by-4.0
8	ARK ASR 0.6B Edge0	4.56%	1.15 B	19	673×	apache-2.0
9	granite speech 4.1 2b ibm-granite	4.62%	2 B	6	547×	apache-2.0
10	MOSS Transcribe Diarize OpenMOSS-Team	4.64%	0.91 B	50	387×	apache-2.0
11	cohere transcribe 03 2026 CohereLabs	4.67%	2 B	14	924×	apache-2.0
12	parakeet tdt 0.6b v2 (fast gpu asr) nvidia	4.67%	0.6 B	—	17K×	cc-by-4.0
13	granite speech 4.1 2b nar ibm-granite	4.68%	2 B	5	2K×	apache-2.0
14	parakeet tdt 0.6b v2 nvidia	4.7%	0.6 B	—	6K×	cc-by-4.0
15	parakeet tdt 0.6b v3 (fast gpu asr) nvidia	4.71%	0.6 B	26	14K×	cc-by-4.0
16	granite speech 5.0 470m turboctc nc ibm-granite	4.84%	0.47 B	—	21K×	cc-by-nc-sa-4.0 non-commercial
17	parakeet tdt 0.6b v3 nvidia	4.84%	0.6 B	26	6K×	cc-by-4.0
18	Voxtral Small 24B 2507 mistralai	4.99%	24 B	8	101×	apache-2.0
19	Phi 4 multimodal instruct microsoft	5.02%	6 B	8	163×	mit
20	granite speech 5.0 470m turboctc ibm-granite	5.03%	0.47 B	—	21K×	apache-2.0
21	Qwen3 ASR 0.6B Qwen	5.05%	0.78 B	52	750×	apache-2.0
22	granite 4.0 1b speech ibm-granite	5.1%	2 B	6	667×	apache-2.0
23	Phonon 2 FermionResearch	5.17%	0.6 B	1	2K×	cc-by-4.0
24	nemotron speech streaming en 0.6b nvidia	5.24%	0.62 B	—	1K×	nvidia-open-model-license
25	granite speech 3.3 8b ibm-granite	5.26%	9 B	5	263×	apache-2.0
26	Zipformer cr ctc transducer XL 290M (fast gpu asr, transducer modified beam search) soundsgoodai	5.29%	0.29 B	—	22K×	cc-by-nc-4.0 non-commercial
27	granite speech 3.3 2b ibm-granite	5.36%	3 B	5	509×	apache-2.0
28	distil large v3.5 distil-whisper	5.4%	0.8 B	1	898×	mit
29	Voxtral Mini 3B 2507 mistralai	5.54%	5 B	8	180×	apache-2.0
30	VibeVoice ASR microsoft	5.58%	8 B	50	218×	mit
31	canary 1b v2 nvidia	5.71%	1 B	25	2K×	cc-by-4.0
32	Fun ASR Nano 2512 FunAudioLLM	5.74%	0.83 B	4	506×	apache-2.0
33	whisper large v3 openai	5.78%	2 B	99	471×	apache-2.0
34	parakeet ctc 1.1b nvidia	5.92%	1.1 B	—	5K×	cc-by-4.0
35	owsm ctc v4 1B espnet	6.28%	1 B	75	783×	cc-by-4.0
36	whisper large v3 turbo openai	6.36%	0.8 B	99	811×	mit
37	Zipformer cr ctc transducer XL 290M (fast gpu asr, ctc greedy search) soundsgoodai	6.38%	0.29 B	—	24K×	cc-by-nc-4.0 non-commercial
38	omniASR LLM 7B v2 facebook	6.4%	7.8 B	1676	130×	apache-2.0
39	Voxtral Mini 4B Realtime 2602 mistralai	6.46%	4 B	13	103×	apache-2.0
40	Audio8 ASR 0.1B Edge0	6.95%	0.324 B	7	726×	apache-2.0
41	niagara 84m batch.en abr-ai	7.46%	0.084 B	1	2K×	applied-brain-research-open-license
42	nemotron 3.5 asr streaming 0.6b nvidia	7.88%	0.64 B	40	1K×	OpenMDW-1.1
43	niagara 38m batch.en abr-ai	8.74%	0.038 B	1	5K×	applied-brain-research-open-license
44	omniASR CTC 7B v2 facebook	9.07%	6.5 B	1676	522×	apache-2.0
45	gemma 4 E4B it google	9.51%	7.94 B	140	161×	apache-2.0
46	niagara 19m batch.en abr-ai	10.41%	0.02 B	1	4K×	applied-brain-research-open-license
47	gemma 4 E2B it google	11.65%	—	140	191×	apache-2.0
48	niagara 9m batch.en abr-ai	13.49%	0.009 B	1	7K×	applied-brain-research-open-license
49	gemma 4 12B it google	20.87%	—	140	106×	apache-2.0

Word error rate: the share of words the model gets wrong, lower is better; average over 8 English test sets (AMI-Cleaned, Earnings22-Cleaned-AA-chunked, Gigaspeech-Cleaned, LS Clean, LS Other, SPGISpeech, Voice Arena Monsoon, Voxpopuli-AA-Cleaned). Speed: how many times faster than real time on the leaderboard’s GPU. Source: Open ASR Leaderboard (Hugging Face), open models only; we recompute the average on the test sets all of them share, so the order can differ slightly from the original page.

Text to speech (audio out)

kokoro 1211 Orpheus TTS 1198 Qwen3 TTS Voice Design 1189 Qwen3 TTS 1125 xtts 1114 Chatterbox 1099 MOSS TTS 1099 IndexTTS 1076 MegaTTS3 1072 parler tts 1058 SparkTTS 1052 fish speech 1 1051

100011001200

Score from blind votes (1000 = average), axis starts at 950. Higher is better. Computed by us from the arena’s public votes with the Bradley-Terry method, like Arena.

What this shows · How to read it

What it shows. The 12 open voice models people preferred most in blind listening tests.

How to read it. Longer bar, preferred more often. 1000 is the average voice; 100 points more means it wins about 64% of the time against an average one.

#	Voice model	Score	±	Votes	Last vote
1	kokoro hexgrad/kokoro	1211	±4	34K	20 Oct 2026
2	Orpheus TTS MohamedRashad/Orpheus-TTS	1198	±5	19K	20 Oct 2026
3	Qwen3 TTS Voice Design Qwen/Qwen3-TTS-Voice-Design	1189	±16	2K	19 Jun 2026
4	Qwen3 TTS Qwen/Qwen3-TTS	1125	±8	8K	20 Oct 2026
5	xtts coqui/xtts	1114	±7	11K	20 Feb 2026
6	Chatterbox ResembleAI/Chatterbox	1099	±10	6K	10 Mar 2026
7	MOSS TTS OpenMOSS-Team/MOSS-TTS	1099	±17	2K	20 Oct 2026
8	IndexTTS IndexTeam/IndexTTS	1076	±16	2K	15 Jul 2025
9	MegaTTS3 ByteDance/MegaTTS3	1072	±14	3K	30 Nov 2025
10	parler tts parler-tts/parler_tts	1058	±7	11K	23 Apr 2025
11	SparkTTS thunnai/SparkTTS	1052	±15	2K	15 Jul 2025
12	fish speech 1 fishaudio/fish-speech-1	1051	±9	6K	8 Apr 2025
13	chatterbox turbo ResembleAI/chatterbox-turbo-demo	1051	±7	10K	19 Oct 2026
14	csm 1b sesame/csm-1b	1050	±18	2K	15 Jul 2025
15	maskgct audio lab Svngoku/maskgct-audio-lab	1042	±12	3K	23 Feb 2025
16	large parler-tts/parler_tts/large	1041	±13	3K	8 Mar 2025
17	openaudio s1 mini fishaudio/openaudio-s1-mini	1037	±16	2K	6 Feb 2026
18	KaniTTS nineninesix/KaniTTS	1036	±29	639	27 Feb 2026
19	llasa 3b tts srinivasbilla/llasa-3b-tts	1031	±11	4K	20 Feb 2026
20	Zonos Steveeeeeeen/Zonos	1028	±21	1K	6 Aug 2025
21	MetaVoice 1B v0.1 mrfakename/MetaVoice-1B-v0.1	1021	±27	706	25 Oct 2024
22	hybrid Steveeeeeeen/Zonos/hybrid	1015	±18	2K	9 Apr 2025
23	llasa 8b tts srinivasbilla/llasa-8b-tts	1015	±23	936	23 Feb 2025
24	E2 F5 TTS mrfakename/E2-F5-TTS	1001	±5	20K	20 Oct 2026
25	MeloTTS mrfakename/MeloTTS	1000	±8	8K	8 Mar 2025
26	parler tts expresso parler-tts/parler-tts-expresso	980	±13	3K	9 Feb 2025
27	GPT SoVITS v2 lj1995/GPT-SoVITS-v2	975	±17	2K	4 Apr 2025
28	style tts 2 Pendrokar/style-tts-2	974	±14	3K	9 Feb 2025
29	multi parler tts PHBJT/multi_parler_tts	973	±28	719	15 Jul 2025
30	kanitts 2 en nineninesix/kanitts-2-en	947	±29	680	10 Mar 2026
31	Llasa 1B finetuned for two speakers HKUST-Audio/Llasa-1B-finetuned-for-two-speakers	930	±25	878	21 Mar 2025
32	EnglishToucan Flux9665/EnglishToucan	896	±26	814	16 Dec 2024
33	OuteTTS 0.3 1B OuteAI/OuteTTS-0.3-1B-Demo	896	±21	1K	25 Jan 2025
34	OpenVoiceV2 myshell-ai/OpenVoiceV2	888	±35	424	25 Oct 2024
35	OpenVoice myshell-ai/OpenVoice	876	±37	378	6 Oct 2024
36	xVASynth TTS Pendrokar/xVASynth-TTS	860	±14	3K	15 Feb 2026
37	HierSpeech TTS LeeSangHoon/HierSpeech_TTS	859	±22	1K	5 Jan 2025
38	WhisperSpeech collabora/WhisperSpeech	844	±22	1K	15 Nov 2024
39	NoDeepMoji Pendrokar/xVASynth-TTS/NoDeepMoji	843	±32	659	4 Apr 2025
40	GPT SoVITS ProPlus lj1995/GPT-SoVITS-ProPlus	840	±12	6K	15 Feb 2026
41	OuteTTS 0.2 500M ameerazam08/OuteTTS-0.2-500M-Demo	840	±20	2K	25 Jan 2025

People hear two voices reading the same text, without names, and pick the better one. Source: TTS Spaces Arena (Pendrokar), 94K votes on open models running on Hugging Face Spaces; paid cloud voices are left out. Models need 300 votes to be ranked. Many votes are from 2024–2025: “last vote” shows how recent each model’s votes are.

4 models with fewer than 300 votes

#	Voice model	Score	±	Votes	Last vote
1	parler tts mini parler-tts/parler_tts_mini	1125	±199	20	2 Aug 2024
2	Pocket TTS NeuralFalcon/Pocket-TTS	1115	±59	142	18 Feb 2026
3	MassivelyMultilingualTTS Flux9665/MassivelyMultilingualTTS	921	±58	153	25 Oct 2024
4	Text to Speech amphion/Text-to-Speech	704	±115	43	11 Apr 2024

Open voices on TTS Arena V2

TTS Arena V2 ranks mostly paid voices; these are the open ones it lists, with its own scores (a different scale from the table above). Source: TTS Arena V2.

Voice model	TTS Arena V2 score	±	Votes	Still in the arena
Tontaube V1 (pre-filter-fix)	1608	±205	4	yes
Chatterbox	1476	±20	2K	yes
Kokoro v1.0	1473	±23	1K	retired
CosyVoice 2.0	1443	±83	58	retired
Maya 1	1406	±36	467	retired
Tontaube V1	1392	±99	51	yes
Fluxions Vui	1364	±92	69	yes
Veena	1360	±34	510	retired

About these rankings

These leaderboards are run by others, not by Arena and not by us. None of them states a license for reusing its results, so we show them with credit and a link to the original. Data updated 20 Oct 2026. Want to compare text models instead? See the free leaderboard or vote in Free Arena.

Share this page:X LinkedIn Reddit WhatsApp