TTS-Model-Comparison-Chart

KO | EN

gitlite — search

#javascript #python #typescript #go #android #machine-learning #macos #nodejs #awesome #kubernetes #vue #security #ios #openai

TTS-Model-Comparison-Chart

(★ 16)

No description available.

TTS-Model-Comparison-Chart 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Local TTS Model Comparison

Personal benchmark of open-source / local text-to-speech models.

**Hardware:** NVIDIA RTX 4090 Laptop GPU  
**Test prompt:** *"The quick brown fox jumped over the lazy dog. She laughed softly and said — well, that was unexpected. [sighs] I suppose we'll just have to try again."*

---

## Model Reviews

Personal notes on each model after hands-on testing. Real Time Factor (RTF) refers to Generation Time divided by Generated Audio Time. I.E: If it takes 60 seconds to generate 30 seconds of audio, its RTF would be 2. The lower the RTF the faster the response.

---

### Bark

Although the main Bark repo doesn't natively support voice cloning itself, there is a 'bark-with-voice-clone' repo that I included with the one click installer. The voice cloning does work, but it isn't the best. The voice does somewhat resemble the cloned one, but changes slightly every generation so it's not the most consistent - almost as if you varied your own voice everytime you talk, but with a different starting voice. Anyways, It can be fun, quirky, and expressive when you get lucky — but oftentimes inconsistent. As far as voice cloning goes - the emotion tags (`[laughter]`, `[sighs]`, etc.) work sometimes and can produce surprising results, but you never quite know what you'll get. Sometimes (with voice cloning) it may skip emotion tags entirely. I will say though, it's kinda impressive it can take a voice (without laughing, sighing, etc) and relatively apply those emotions to a voice that wasn't trained on it. Voice cloning is limited; the `.npz` approach works but the clone just barely sounds close to the reference. It can also sometimes switch voices too, for example, if you clone a female voice, you may get a male voice, although rare in my testing, but still noteworthy to mention. Otherwise, it isn't really that expressive. It sounds almost half robotic and half human. It does work a little better without voice cloning though, if you don't mind the presest voices, which there's quite a few of. Best used for creative/character TTS where some randomness is acceptable, not for production, narration work, or voice assistants. It's also not the fastest, nor the slowest either with 1.35 RTF. Note: Although the official repository seems to use brackets for emotions tags, I've found that parenthesis seem to be better. Using the brackets increases the chances it'll sound less like the voice clone and vary the voice more. The voice cloning also increases the RTF from 1.35 to about 1.43 as well. Longer sentences can increase the likelihood that the voice deviates, but there is a --rolling flag that I added in you can use - meaning - it should use the context of the voice prior to the chunking, but this doesn't always work the best. There is also a pre-existing --seed flag you can use as well which allows you to use the seed you liked for reproducability to mitigate these issues. Voices are more consistent if don't use voice cloning. It's supposed to support music and singing, which would increase it's emotion score, but I've found it to not to not work nearly as well or even close to the demo in the repo.

---

### Dia — robertagee

This is the best emotion-tag model I've tested locally. Twenty-one tags that produce real audio events — laughs, gasps, sighs, etc.  — not just tone shifts. Some of the tags don't work that well with voice cloning, however. The one's that don't work as well have already been tested and are listed in the main one click installer repo I created. Voice cloning via a reference transcript also works well. It can be slow at ~2.2× RTF, but the expressiveness payoff is worth it for short clips as well as the decent clone quality. Consistency is pretty solid as well.. It's not the best, but it does sound like the cloned voice, and keeps it's likeness throughout generations (although does vary slightly). If you want it to be faster, you can use the --compile flag which uses torch.compile to optimize. This brings RTF down to a whopping .77! Faster than real time! Although that doesn't mean it responds right away, it just means that the audio takes less time to generate than the clip is long. The expressiveness is very advanced. Almost too advanced. By that, I mean, it almost takes away it's human likeness because it's too expressive at times. But if you want a fun, expressive voice, then this is a great model to use. If this model was paired with the speed of Kokoro it'd be very hard to beat. Then if you added F5's voice cloning ability, and it would become SOTA in my opinion. In my opinion, Dia is a better more refined (and more importantly - usable) "version" of BarkTTS.

---

### Dia2

Actually way worse than Dia. Not sure why. I even installed Triton and it uses Cuda-Graph too. I messed around a long time with this one to see if Claude Opus 4.5 could optimize it, but it's RTF was about 3.7 at best. Significantly less reliable clone quality in my tests too.

---

### Fun-CosyVoice3

This is a very interesting & unique model. While it doesn't have as many emotion tags (like Dia or Bark), the one's it does have are decent. But not only does it have emotion tags like laughter, or cough, quick cough, etc. it also has emotion dictating, where you can give it a tag like "anger", "whisper", "fast" or "slow", and it has an "instruct" ability - where you can pass in and tell it how you want it to sound and it'll make the whole sentence sound that way, like it's yelling, talking quietly, etc. I will say for "instruct" it only affects about 25% as much as you'd probably hope it would, and similary for the "emotion" tags, but it does make a slight difference. There hasn't been another TTS I've tested so far with this ability. As mentioned - I will say it doesn't work the best though. The "anger" one in particular doesn't always sound like it's yelling, but sometimes more like it's annoyed. It can vary, but generally gets it close to the desired outcome. The voice is pretty consistent throughout several generations as well, which is nice. So far, it only varies very slightly, so it should always sound like the same speaker. As far as the voice clone quality, it's pretty close. Definitely recognizable from the original cloned speaker. While it can be expressive, and is good at making the cloned original speaker sounding like it's laughing, angry etc. in it's original cloned voice (which is really impressive actually since it wasn't trained on it and only cloned a few seconds in a "one shot" clone), it does sound bland and robotic at times or unnatural. Not as natural as human. But still, would recommend this model for character creation and alike - if you don't mind a few generations because it just may take a few generations to get it right. The best thing about this model is it's the fastest emotion capable and voice cloning model yet! With an RTF around .5 to .6 ~ about half real time! If you're generating longer audio, and need emotion and specific voices this is a very good choice for that reason. Because of it's lower RTF it also makes it a contender for a voice assitant as well, although there will be some delay - especially the longer the audio since it's about .6 RTF. A Note: Sometimes, especially if you're giving it a bigger paragraph, it will kind of sound like random half english words/sentences at times. Where it sounds like the language, but isn't understandable unfortunately. for this reason I wouldn't suggest using it for an assitant, or scripting, etc. because you may have to redo the generation several times to get longer pieces of text correct.

---

### Higgs Audio v2

This is another unique model. For emotion control, it doesn't use emotion tags, but instead infers expression from context. You can also dictate emotion via a --scene flag and let it be more expressive with --temp flag. Although keep in mind with higher temperates, you may get less consistent results. This model also support cloning as well, and it's pretty decent! For this model 24 GB VRAM is recommended but you can get away with 16 GB of VRAM as well. RTF can be around .77 or so if using the --fast-cache flag, otherwise it's closer to 1.42. This is by far the heaviest VRAM model yet, using 14.2 GB without the --fast-cache flag and 16 GB+ on my 4090 Laptop with --fast-cache flag. As mentioned before this model has several options for voice control. The first option is voice cloning. It's not as good as F5, but it's pretty good and keeps the voice somewhat consistent at lower tempteratures. It has a "--scence" flag which you can describe the voice you want. For example you can say things like "a young nasely man speaking", etc. It also has preset voices as well you can use. And these are additive, so you can combine them together with voice cloning. Use the --temp flag to control expressiveness (although you do risk more random and inconsistent generations this way). Unfortunately, the emotion tags only seem to work rarely, so this isn't a model to use if you want them. I'd say 1 out of every 20 generations for when emotion tags work on average. Also one thing to note, if you don't have the recommended VRAM and you use this model, after a few generations it will take longer and longer to load in interactive mode. This model would be a good choice if you need specific control of a voice and fine tuning, and don't mind playing around with it awhile to get the perfect voice and don't mind several generations. It can generate artificts as well, with one generation having a random beeping sound for over a minute of audio. So beware of things like that happening. Lastly, pausing and speed at which talks on this model is very realistic as well. Keep in mind when trying to voice cloning, the scene and temperature tags will have less effect than if using without. I would not recommend this for a voice assitant while consistent, because of how much VRAM it uses when there are models that offer a similar enough quality at a fraction of the VRAM usage. A temp of .8 seems to do well for an in between of consistency and expressiveness for voice cloning. It also does seem to cut words off at very end of the text as well. When not using voice cloning the voices can vary quite a bit. For example, it may sound male or female, but vary quite a bit. Temperature may help with this though.

---

### Orpheus

This is a decent model with emotion tags like `<laugh>`, `<sigh>`, etc. The tags produce actual audio events, not just prosody shifts. Voice consistency is good and realism is decent as well. It sounds fairly human and realistic. I will say the pausing on this model is really good - probably one of the best and more realistic I've tested. It does take a long time to generate though at about 2 RTF. But it gets realistic expressiveness down pretty well. It sounds like a human talking neutrally- similar to ChatterBox. Voice cloning is now supported via `--ref-audio` / `--ref-text`, which switches to the pretrained base model automatically. **Important caveat:** emotion tags do not work with voice cloning — the pretrained model was never fine-tuned to respond to them. So it's emotion tags OR voice cloning, not both at the same time. Unfortunately when using the voice cloning, it can repeat sentences randomly so I would not recommend the voice cloning for Oprheus. It also takes even longer to generate as well at about 3 RTF for longer paragraphs - one of the worst I've tested. Because it's voice cloning isn't the best quality, and it doesn't have a worthy edge on any of the others voice cloning abilities, there isn't a reason to pick it over TTS models. Also when using voice cloning it sounds a lot more bland as well, reducing its expressability. If you don't mind preset voices, then this is good choice when you want reliable, realistic delivery with some emotion control on a built-in voice. For voice cloning it also often gets some weird non-language sounding sounds and weird artifacts - even for the preset voices, just not as often. Especially the longer your text is. But not as significant as CosyVoice. For this reason, I wouldn't recommend it for an assistant, scripting, etc unless several generations are okay. Quality does improve for shorter sentences though. This is also one of the heavier models at about 9 GB of VRAM.

---

### Chatterbox

Shockingly clean output — perfect consistency (5/5) and no trailing noise and negligible cutoffs. No emotion tags, but the `exaggeration=` parameter (0.0–2.0) pulls some range out of it. Voice clone quality is one of the best I've tested at 5/5 - I'd say it rivals Elevan Labs and F5-TTS. For this test I'm using the fast rsxdalv variant which brings it down to speedy 0.5× RTF! This is my go-to model when quality and reliability matter more than expressiveness and emotion tags. I will say though, it **can** get a bit of a british accent at times - not sure why. But it's not that different from the original speaker, just noticeable. This is one of the more realistic sounding ones - if you want it to sound human, and not robotic, this is a very good choice. The best way I can describe this model - is that it sounds like an unexpressive human. For example, it won't sound sad, angry, happy etc. but more more neutral. Almost like a professor teaching a class.

---

### IndexTTS2

This model excels at the ability to pull range out of emotion expressiveness - even when voice cloning. It beats most other TTS models by a large factor in this regard. It has ability to sound happy, angry, sad, afraid, disgusted, melancholic, suprised, and calm. While it has the ability to pull range of emotion, it's important to note it's not pefect. When testing with the same phrase: For example, for some reason happy, afraid, and sad sound very similar, in the way that they're exaggerated. When using angry it sounds more excited or surprised in the way it shapes the voice. When using the disgusted option it sounds more annoyed than digusted. Melancholic does sound pretty sad though. The surprised option sounds like more like a monotone and sarcastic surprised. Similarly, the calm option sounds a little snarky or sarcastic as well. I'm not sure if it's a completely fair way to compare though, because our brains aren't used to hearing the same phrase in different tones. But thought it'd be a way to consistently see how the tone changes without the variable of different of text. To go back to voice cloning - the cloned voice does sound very close to the refence clip. Not as good as Eleven Labs but very close! Furthermore, it does a surprisingly good job at taking the cloned voice and applying the prior mentioned emotions/tones to the voice. Meaning - it still sounds like the cloned reference, even when applying the emotions/tones to the voice. Another thing this model excels at is keeping the prosody and inflection of the cloned voice as well. I will say, at least with voice cloning, it does have many problems with non word text i.e: apostrophes, hypens, etc. It will say things like "don T" instead of "don't" and so forth. It's a very heavy model when it comes to VRAM - at 14 GB of VRAM. So you will need at least a 16 GB VRAM GPU as well. For the reasons of non word text issues regarding voice cloning, I would not recommend this model, unless you're willing to do work arounds for scripting. Same regards as a voice assitant. It's unfortunate, because otherwise this model is very expressive. Especially with how well it clones the original speaker.

---

### F5-TTS

Fastest model in the lineup by a wide margin (RTF 0.33). Voice cloning is excellent — easily the best clone quality I've tested (4.5/5). No emotion tags, purely neutral TTS, but the voice reproduction is remarkably clean. Great pick when you need fast, high-quality cloned speech without any expressiveness requirements.

---

### VibeVoice

Interesting Microsoft research model — runs at nearly real-time (RTF ~0.93) and cloning works. Output is clean with minimal trailing issues. Not particularly expressive, but a solid all-rounder for neutral cloned speech. Marked as research-only so production use is uncertain.

---

### Coqui TTS

Very slow (RTF 4.33) and inconsistent quality. Cloning exists but the results aren't competitive with newer models. Fine for light local use but hard to recommend over anything else on this list given the speed.

---

### OpenVoice

Emotion control is weak (1.0/5) and clone quality is mediocre. Has trailing noise issues. Hasn't aged well — superseded by better cloning models. Keeping it benchmarked for historical reference.

---

### Kyutai

Excellent consistency (5.0/5) and very expressive (4.0/5). Runs at real-time (RTF 1.0). No voice cloning in the base model. Setup was moderately involved. One of the more impressive sounding models — wish it had clone support.

---

### Kyutai-Mimi

Same consistency as Kyutai but expressiveness drops to 2.0/5 in the Mimi variant. Clone support added but clone quality is poor (2.0/5). Notably harder to set up (difficulty 4.0). Mostly superseded by base Kyutai.

---

### Fish Speech S2 Pro

High VRAM (14–15 GB) but the voice cloning is very good and it supports a large number of speakers. Emotion tags are present but produce only subtle prosody changes. Worth running if you have the VRAM headroom and want a polished cloning experience.

---

### Qwen3-TTS

Haven't benchmarked yet. NL-instruction emotion control — you describe the emotion in a separate instruction string. Flexible and interesting approach. Runs at 3–8 GB VRAM depending on the variant. Voice cloning supported.

---

### MegaTTS 3

Haven't benchmarked yet. No emotion tag support — prosody entirely from voice clone. High clone quality expected based on claims. 8–16 GB VRAM range.

---

### Kokoro, XTTS, Sesame CSM, GPT-SoVITS, ChatTTS, FireRedTTS, CosyVoice2, MaskGCT, Zonos, AllTalk TTS

Not yet tested.

---

## Speed — Real-Time Factor (RTF)

> **RTF = Generation Time ÷ Audio Length**  
> RTF < 1.0 = faster than real-time | RTF = 1.0 = real-time | RTF > 1.0 = slower

| Model | 12s Time | 12s RTF | 30s Time | 30s RTF | 60s Time | 60s RTF | VRAM Usage |
|---|---|---|---|---|---|---|---|
| **Bark** | 16.4s | 1.33 | 42s | 1.35 | 81.4s | 1.37s | 5.8 GB |
| **Bark (Voice Clone)** | 16.9s | 1.41 | 42.9s | 1.43 | 85.2s | 1.42s | 6.0 GB |
| **Dia - robertagee** | 27.6s | 2.2 | 66.1s | 2.2 | — | — | — |
| **Dia - robertagee (--compile)** | 27.6s | 2.2 | 66.1s | 2.2 | — | — | — |
| **Dia (Voice Clone) - robertagee** | 27.6s | 2.2 | 64.9s | 2.16 | 126 | 2.1 | 7.2 GB |
| **Dia - (Voice Clone) - robertagee (--compile)** | 9.4s | **0.77** | 23.1s | **0.77** | 47s | **0.78** | 6.2 GB |
| **Dia2** | 44.4 | 3.7 | — | — | — | — | — |
| **Chatterbox** | 24s | 1.67 | — | — | — | — | — |
| **Chatterbox - rsxdalv fast** | 6s | **.5** | 15.3s | **.51** | 29.1 | **.48** | 5.1 GB |
| **Chatterbox (Voice Clone) - rsxdalv fast** | 6s | **.5** | 15.3s | **.51** | 29.1 | **.48** | 5.7 GB |
| **CosyVoice3** | 7.5 | **.63** | 17.1 | **.62** | 31s | **.52** | 4.5 GB |
| **Higgs Audio V2** | 17.4s | 1.4 | 41.7 | 1.39 | 85.2s | 1.42| 14.2 GB |
| **Higgs Audio V2 (--fast-cache)** | 11s | **.77** | 23.8s | **.79** | 52s | **.87** | 16 GB+ |
| **F5-TTS** | 4s | 0.33 | — | — | — | — | — |
| **Orpheus - CLI** | 22s | 1.8 | 66.4s | 2.2 | 146s | 2.4 | 9.1 GB |
| **Orpheus (Voice Clone) - CLI** | 23.5s | 1.96 | 76s | 2.5 | 191s | 3.2 | 9.3 GB |
| **Orpheus** | 12s | 1.00 | — | — | — | — | — |
| **IndexTTS2** | 9.6 | .8 | 25s | .83 | 50.4s | .83 | 14 GB |
| **VibeVoice** | 11s | .92 | 28.2s | .94 | 57s | .95 | 1.4 GB |
| **Coqui TTS** | 52s | 4.33 | — | — | — | — | — |
| **OpenVoice** | 19s | 1.58 | — | — | — | — | — |
| **Kyutai** | 12s | 1.00 | — | — | — | — | — |
| **Kyutai-Mimi** | 12s | 1.00 | — | — | — | — | — |
| **Higgs** | — | — | — | — | — | — | — |
| **Kokoro** | — | — | — | — | — | — | — |
| **XTTS** | — | — | — | — | — | — | — |
| **Sesame CSM** | — | — | — | — | — | — | — |
| **GPT-SoVITS** | — | — | — | — | — | — | — |
| **ChatTTS** | — | — | — | — | — | — | — |
| **Fish Speech** | — | — | — | — | — | — | — |
| **FireRedTTS** | — | — | — | — | — | — | — |
| **CosyVoice2** | — | — | — | — | — | — | — |
| **MaskGCT** | — | — | — | — | — | — | — |
| **Zonos** | — | — | — | — | — | — | — |
| **AllTalk TTS** | — | — | — | — | — | — | — |
| **Qwen3-TTS** | — | — | — | — | — | — | — |
| **MegaTTS 3** | — | — | — | — | — | — | — |

> `—` = not yet benchmarked. `❌` = could not get running.

---

## Quality Ratings

> All scores out of 5. Higher is better.  
> **Difficulty** = ease of install: 1 = one-click easy, 5 = failed to get running.

| Model | Emotion Tags | Emotions | Expressiveness | Consistency | No Artifacts | No Trailing | No Cutoff | Realism | Voice Cloning | Clone Quality | Difficulty |
|---|---|---|---|---|---|---|---|---|---|---|---|
| **Bark** | Yes | 3.5 | 2.5 | 4.0 | 3.5 | 4.5 | 3.5 | 2.5 | No | — | 1.0 |
| **Bark (Voice Clone)** | Yes | 2.5 | 2.5 | 2.0 | 3.5 | 4.5 | 3.5 | 2.5 | Yes | 1.5 | 1.0 |
| **Dia - robertagee** | Yes | 4.0 | 3.0 | 2.0 | — | 4.0 | 3.0 | 3.5 | Yes | 2.5 | 1.0 |
| **Dia (Voice Clone) - robertagee** | Yes | 4.0 | 3.5 | 4.0 | 3.5 | 4.5 | 3.5 | 3.5 | Yes | 3.5 | 1.0 |
| **Dia (Voice Clone) - robertagee (--compile)** | Yes | 4.0 | 3.5 | 4.0 | 3.5 | 4.5 | 3.5 | 3.5 | Yes | 3.5 | 1.0 |
| **Dia2** | Yes | 1.0 | 2.0 | 2.0 | 0.5 | 2.0 | 3.0 | 0.5 | Yes | 0.0 | 4.0 |
| **Chatterbox** | No | — | 3.0 | 5.0 | — | 5.0 | — | — | Yes | 5.0 | 1.5 |
| **Chatterbox - rsxdalv fast** | No | — | 4.0 | 5.0 | 4.5 | 5.0 | 5.0 | 4.5 | No | - | 2 |
| **Chatterbox (Voice Clone) - rsxdalv fast** | No | — | 4.0 | 5.0 | 4.0 | 5.0 | 5.0 | 4.0 | Yes | 5.0 | 1.5 |
| **Fun-CosyVoice3** | Yes | 3.5 | 3.5 | 4.0 | 2.0 | 4.5 | 4.5 | 2.0 | Yes | 4.0 | 1.0 |
| **Higgs Audio v2** | No | — | 4.5 | 4.0 | 3.5 | 4.0 | 3.5 | 3.5 | Yes | 4.0 | 2.0 |
| **Higgs Audio v2 (--fast-cache)** | No | — | 4.5 | 4.0 | 3.5 | 4.0 | 3.5 | 3.5 | Yes | 4.0 | 2.0 |
| **F5-TTS** | No | — | 3.0 | 4.5 | — | 5.0 | — | — | Yes | 4.5 | 2.0 |
| **Orpheus - CLI** | Yes | 2.5 | 3.5 | 4.0 | 2.0 | 2.0 | 2.0 | 2.5 | No | — | 1.0 |
| **Orpheus (Voice Clone) - CLI** | No | — | 2.0 | 2.0 | 1.0 | 1.0 | 2.5 | 1.0 | Yes | 2.0 | 1.0 |
| **Orpheus** | Yes | 2.5 | 2.5 | 4.0 | — | 3.0 | — | — | No | — | 2.0 |
| **VibeVoice** | No | 0 | 2.0 | 3.5 | — | 4.5 | 4.5 | 2.0 | Yes | 3.5 | 1.0 |
| **IndexTTS2** | No | — | 4.5 | 4.5 | 3.5 | 4.5 | 3.5 | 4.0 | Yes | 4.5 | 1.0 |
| **Coqui TTS** | — | 2.5 | 3.0 | 1.0 | — | 2.0 | — | — | Yes | 2.5 | 3.0 |
| **OpenVoice** | — | 1.0 | 2.0 | 3.5 | — | 2.0 | — | — | Yes | 2.0 | 2.0 |
| **Kyutai** | — | — | 4.0 | 5.0 | — | 5.0 | — | — | No | — | 3.0 |
| **Kyutai-Mimi** | — | — | 2.0 | 5.0 | — | 5.0 | — | — | Yes | 2.0 | 4.0 |
| **Higgs** | — | — | — | — | — | — | — | — | — | — | — |
| **Kokoro** | — | — | — | — | — | — | — | — | — | — | — |
| **XTTS** | — | — | — | — | — | — | — | — | — | — | — |
| **Sesame CSM** | — | — | — | — | — | — | — | — | — | — | 5.0 |
| **GPT-SoVITS** | — | — | — | — | — | — | — | — | — | — | 5.0 |
| **ChatTTS** | — | — | — | — | — | — | — | — | — | — | — |
| **Fish Speech** | — | — | — | — | — | — | — | — | — | — | — |
| **FireRedTTS** | — | — | — | — | — | — | — | — | — | — | — |
| **CosyVoice2** | — | — | — | — | — | — | — | — | — | — | — |
| **MaskGCT** | — | — | — | — | — | — | — | — | — | — | — |
| **Zonos** | — | — | — | — | — | — | — | — | — | — | — |
| **AllTalk TTS** | — | — | — | — | — | — | — | — | — | — | — |
| **Qwen3-TTS** | — | — | — | — | — | — | — | — | Yes | — | — |
| **MegaTTS 3** | — | — | — | — | — | — | — | — | Yes | — | — |

**Column definitions:**
- **Emotion Tags** — supports emotion/tone markup tags in input text (e.g. `[laughs]`, `(whispers)`, `<happy>`)
- **Emotions** — how well and consistent it can express different emotional tones via tags (happy, sad, angry, etc.)
- **Expressiveness** — how natural and varied the delivery sounds
- **Consistency** — does the voice stay stable and consistent across multiple generations
- **No Artifacts** — absence of digital glitches, clicks, static, distortion, or unnatural audio discontinuities mid-speech
- **No Trailing** — does audio end cleanly without silence/noise or random noise after the speech
- **No Cutoff** — does speech complete fully without words being cut off at the end
- **Realism** — how human and lifelike the voice sounds (naturalness of prosody, breath, tone)
- **Voice Cloning** — supports cloning a voice from a reference audio file
- **Clone Quality** — how well it reproduces a reference voice (if supported)
- **Difficulty** — 1 = runs out of the box, 5 = could not get working

---

## Audio Samples

One male and one female sample per model, generated from the same test prompt.  
Click a link → GitHub opens the file page with an inline audio player.

| Model | Female | Male |
|---|---|---|
| **Bark** | [▶ female](samples/bark/female.wav) (en_speaker_9) | [▶ male](samples/bark/male.wav) (en_speaker_6) |
| **Dia** | [▶ female](samples/dia/female.wav) | [▶ male](samples/dia/male.wav) |
| **Dia2** | — | — |
| **F5-TTS** | [▶ female](samples/f5_tts/female.wav) | [▶ male](samples/f5_tts/male.wav) |
| **Orpheus** | [▶ female](samples/orpheus/female.wav) | [▶ male](samples/orpheus/male.wav) |
| **Chatterbox** | [▶ female](samples/chatterbox/female.wav) | [▶ male](samples/chatterbox/male.wav) |
| **Coqui TTS** | [▶ female](samples/coqui/female.wav) | [▶ male](samples/coqui/male.wav) |
| **OpenVoice** | [▶ female](samples/openvoice/female.wav) | [▶ male](samples/openvoice/male.wav) |
| **Kyutai** | [▶ female](samples/kyutai/female.wav) | [▶ male](samples/kyutai/male.wav) |
| **Kyutai-Mimi** | [▶ female](samples/kyutai_mimi/female.wav) | [▶ male](samples/kyutai_mimi/male.wav) |
| **Higgs** | — | — |
| **Kokoro** | — | — |
| **XTTS** | — | — |
| **Sesame CSM** | — | — |
| **GPT-SoVITS** | — | — |
| **ChatTTS** | — | — |
| **IndexTTS2** | — | — |
| **Fish Speech** | — | — |
| **FireRedTTS** | — | — |
| **CosyVoice2** | — | — |
| **MaskGCT** | — | — |
| **Zonos** | — | — |
| **AllTalk TTS** | — | — |
| **Qwen3-TTS** | — | — |
| **MegaTTS 3** | — | — |
| **VibeVoice** | — | — |

---

## Model Repos

Each model lives in its own subfolder with a `setup.ps1` one-click installer and its own `venv/`.

| Model | Status | GitHub |
|---|---|---|
| Bark | ✅ Ready | [Bark_TTS_CLI_Local](https://github.com/mirfahimanwar/Bark_TTS_CLI_Local) |
| Dia | ✅ Ready | [Dia-TTS-CLI-Local](https://github.com/mirfahimanwar/Dia-TTS-CLI-Local) |
| Dia2 | ✅ Ready | [Dia-2-TTS-CLI-Local](https://github.com/mirfahimanwar/Dia-2-TTS-CLI-Local) |
| F5-TTS | — | — |
| Orpheus | ✅ Ready (requires HF login) | [Orpheus-TTS-CLI-Local](https://github.com/mirfahimanwar/Orpheus-TTS-CLI-Local) |
| Chatterbox | ✅ Ready | [Chatterbox-TTS-CLI-Local](https://github.com/mirfahimanwar/Chatterbox-TTS-CLI-Local) |
| VibeVoice | ✅ Ready (research only) | — |
| Coqui TTS | — | — |
| OpenVoice | — | — |
| Kyutai | — | — |
| Kyutai-Mimi | — | — |
| Higgs | ✅ Ready | — |
| Kokoro | — | — |
| XTTS | — | — |
| Sesame CSM | — | — |
| GPT-SoVITS | — | — |
| ChatTTS | — | — |
| IndexTTS2 | ✅ Ready | — |
| Fish Speech S2 Pro | ✅ Ready | — |
| FireRedTTS | — | — |
| CosyVoice2 | — | — |
| MaskGCT | — | — |
| Zonos | — | — |
| AllTalk TTS | — | — |
| Qwen3-TTS | — | — |
| MegaTTS 3 | — | — |

**Community Recommended — Not Yet Benchmarked:**

| Model | Notes |
|---|---|
| MOSS-TTS | OpenMOSS-Team/MOSS-TTS — open weights, emotion control |
| HumeAI/tada | Cloud API only (requires account) |
| SoulX | Soul-AILab/SoulX — expressive, HuggingFace weights |
| ~~VibeVoice~~ | ~~microsoft/VibeVoice~~ — now in [VibeVoiceTTS/](VibeVoiceTTS/) |
| Neuphonic | neuphonic/neutts — cloud API only |
| Supertone-2 | Supertone/supertonic-2 |
| maya-research | Community recommendation, limited info |
| kittenTTS | Community recommendation, limited info |

---

## Notes

- All benchmarks run on **NVIDIA RTX 4090 Laptop GPU** with models loaded fresh (no warm-up cache)
- RTF measured as wall-clock time from command submission to wav file written
- Quality scores are subjective — tested on the same 4–5 sentences per model
- `—` = not yet tested | `❌` = could not get running after reasonable effort
- 30s and 60s RTF columns will be filled in as each model is benchmarked

---

## Expressive TTS — Emotion Tag Index

Which models support inline emotion tags that produce **real audio events** (not just prosody changes).

| Model | Emotion Tags | Tag Format | Sample Tags | VRAM | Voice Cloning |
|---|---|---|---|---|---|
| **Dia / Dia2** | ✅ 21 tags | `(laughs)` | `(laughs)` `(cries)` `(gasps)` `(sighs)` `(growls)` `(groans)` `(screams)` | 4–5 GB | ✅ |
| **Orpheus** | ✅ 8 tags (finetune only) | `<laugh>` | `<laugh>` `<sigh>` `<gasp>` `<groan>` `<cough>` `<sniffle>` `<yawn>` `<chuckle>` | 6–8 GB | ✅ (tags disabled when cloning) |
| **Bark** | ⚠️ limited | `[laughter]` | `[laughter]` `[sighs]` `[clears throat]` | 6–8 GB | ❌ |
| **Fish Speech S2 Pro** | ⚠️ subtle | `[sighs]` | Tags present but produce mild prosody changes only | 14–15 GB | ✅ |
| **Higgs Audio v2** | ✅ emergent | none needed | Pass `instruct=` or embed emotion in text | 24 GB | ✅ |
| **IndexTTS2** | ✅ multi-modal | none needed | `--emo-audio FILE` · `--emo-text "angry"` · `--emo-vector 0,0.8,0,...` | 8–12 GB | ✅ |
| **Qwen3-TTS** | ✅ NL instructions | none needed | `instruct="speak with excitement"` | 3–8 GB | ✅ |
| **MegaTTS 3** | ❌ | none | No tag support — prosody via voice clone only | 8–16 GB | ✅ |
| **Chatterbox** | ❌ | none | Emotion via `exaggeration=` parameter (0.0–2.0) | 4–6 GB | ✅ |

### Best pick by use case

| Goal | Best Model | Why |
|---|---|---|
| Most emotion tags + local | **Dia2** | 21 tags, 4.4 GB, runs well on low VRAM |
| Laugh/cry/gasp audibly | **Orpheus** | Tags produce actual audio events, not just tone |
| No tags, emergent emotion | **Higgs Audio v2** | State-of-the-art EmergentTTS benchmark, 24 GB |
| NL-based emotion control | **Qwen3-TTS** | Natural language instruction, 3–8 GB |
| Voice clone + emotion | **Dia2** | Best combo of tags + clone quality at low VRAM |

### Dia tag list (21 total)

`(laughs)` `(chuckles)` `(giggles)` `(cries)` `(sobs)` `(sighs)` `(gasps)` `(screams)`  
`(moans)` `(groans)` `(growls)` `(sniffs)` `(yawns)` `(coughs)` `(clears throat)`  
`(whistles)` `(mumbles)` `(stutters)` `(hums)` `(shouts)` `(whispers)`

### Orpheus tag list (8 total)

`<laugh>` `<chuckle>` `<sigh>` `<cough>` `<sniffle>` `<groan>` `<yawn>` `<gasp>`

TTS-Model-Comparison-Chart

TTS-Model-Comparison-Chart 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

TTS-Model-Comparison-Chart

TTS-Model-Comparison-Chart 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README