If you want to run open-source AI voice models locally, the hard part is not the installation guide. It is that most of the famous checkpoints cannot legally be used for paid work, and the ones that can are not the ones the community keeps recommending. XTTS-v2 has been downloaded from Hugging Face more than six million times, and its weights are non-commercial. F5-TTS’s code is MIT but its checkpoints are CC-BY-NC because of the training data.
I went through eighteen open-weight speech models on 16 October 2026 and checked the actual licence file for each, plus every current price page for the GPU hosts. Eleven have no blanket commercial licence. That is the single most useful thing in this guide, and it is what almost every other “best local TTS models” list omits.
What can you actually use commercially?
For paid work in October 2026, stick to models whose weights are Apache-2.0 or MIT: Kokoro-82M, Chatterbox, VoxCPM2, Qwen3-TTS, CosyVoice2 and Fun-CosyVoice3, and OpenVoice V2. Everything else needs a legal read before it touches a revenue stream.
Two conditional cases are worth knowing. Fish Audio replaced its old non-commercial terms with the Fish Audio Community License, version 1.0, dated 13 October 2026, which grants commercial rights while your annual revenue is under USD 500,000 and bans offering the model as a hosted TTS service regardless of revenue. IndexTTS 2.5 runs on the bilibili Model Use License Agreement, commercial below 100 million monthly active users and RMB 1 billion in annual revenue.
And the one everybody learns the hard way:
“This license allows only non-commercial use of a machine learning model and its outputs.”: the first line of the Coqui Public Model License 1.0.0, the licence shipped alongside the XTTS-v2 weights
Reading the licence, memory and speed tables
Licence terms come from repository files and Hugging Face model metadata checked on 16 October 2026. Code and weights licences are listed separately where they differ.
“Weights on Hub” refers to downloadable file sizes, not peak VRAM. Hardware requirements and real-time factors come from project maintainers on named hardware. The models were not run for this comparison, so these figures do not establish a hands-on audio-quality ranking. Hosting prices are published list rates.
The licence problem nobody puts in the comparison table
Open-weight does not mean open-source, and neither one means you can sell the output. The most common trap is reading the GitHub repository’s licence and assuming it covers the checkpoint.
| Model | Code licence | Weights licence | Commercial use? |
|---|---|---|---|
| XTTS-v2 (Coqui) | MPL-2.0 | Coqui Public Model License 1.0.0 | No: non-commercial only, including outputs |
| F5-TTS / E2-TTS | MIT | CC-BY-NC-4.0 | No: NC clause, because of the Emilia training data |
| Fish Audio S2-Pro | Fish Audio Community License v1.0 | Fish Audio Research License | No: separate commercial licence required |
| Higgs Audio v3 (Boson AI) | : | Boson Higgs Audio v3 Research and Non-Commercial License | No |
| Spark-TTS | Apache-2.0 | CC-BY-NC-SA-4.0 | No: NC-SA on the weights |
| ChatTTS | AGPL-3.0 | Model card says academic purposes only | No, and AGPL-3.0 is viral on the code |
| Fish Audio Drama 3 | : | Fish Audio Community License v1.0 | Conditional: under USD 500,000 revenue, no hosted service |
| IndexTTS 2.5 | ( | bilibili Model Use License Agreement | Conditional) under 100M MAU and RMB 1bn revenue |
| Piper (piper1-gpl) | GPL-3.0 | Varies per voice | Conditional: copyleft on the engine |
| Seed-VC | : | GPL-3.0 per Hub metadata | Conditional: copyleft |
| so-vits-svc | AGPL-3.0 (svc-develop-team fork) | ( | Conditional) AGPL |
| StyleTTS 2 | MIT | Disclosure-and-consent terms on the pretrained models | Conditional: must announce synthesis unless the speaker consented |
| Kokoro-82M | : | Apache-2.0 | Yes |
| Chatterbox | MIT | MIT | Yes: output is watermarked |
| VoxCPM2 | : | Apache-2.0 | Yes |
| Qwen3-TTS | Apache-2.0 | Apache-2.0 | Yes |
| CosyVoice2 / Fun-CosyVoice3 | Apache-2.0 | Apache-2.0 | Yes |
| Zonos v0.1 | Apache-2.0 | Apache-2.0 | Yes |
| Dia 1.6B | Apache-2.0 | Apache-2.0 | Yes: model disclaimer restricts misuse |
| OpenVoice V2 | MIT | MIT | Yes |
| VibeVoice 1.5B | MIT | MIT | Yes |
| RVC WebUI | MIT | MIT (code) | Yes for the code |
Four details in that table deserve emphasis.
The XTTS trap. The Coqui TTS toolkit is MPL-2.0 and the model card even links the CPML as a separate document. Only the CPML governs your output. It defines non-commercial use as anything where you “receive any direct or indirect payment arising from the use of the model or its output”, and explicitly excludes use by commercial entities even for internal research.
The F5-TTS split. The repository’s own licence section says the code is MIT and the pre-trained models are CC-BY-NC “due to the training data Emilia, which is an in-the-wild dataset”. The Hub metadata agrees.
Fish Audio moved. The current Fish Speech repository LICENSE is the Fish Audio Community License v1.0, last updated 13 October 2026. Section III(a) grants commercial use “only while Your Annual Revenue is less than USD 500,000”, measured across you and your affiliates. Section III(c) carves out Hosted Services entirely: build an API where third parties generate speech from text or voice samples and you need a separate written licence regardless of revenue. The README adds that S2-Pro weights specifically “remain under the Fish Audio Research License”. If a blog post tells you Fish Speech is either fully open or fully closed, it predates this licence.
Piper changed hands and changed licence. The original Rhasspy Piper is MIT but archived. The Open Home Foundation successor, piper1-gpl, is GPL-3.0 and was last pushed on 15 October 2026. GPL on an engine you distribute is a real obligation; the voice files carry their own terms.
VRAM: the real ceiling when you run open-source AI voice models locally
Model quality is what people argue about. Available VRAM is what decides whether it runs. The table separates downloadable weight size from the maintainers’ hardware requirements.
| Model | Weights on Hub (16 Oct 2026) | Vendor GPU statement | My pick for a card |
|---|---|---|---|
| Kokoro-82M | 0.36 GB | none published | 4 GB, or CPU |
| OpenVoice V2 | 0.13 GB | none published | 4 GB |
| Fish Speech 1.5 | 1.46 GB | none published | 8 GB |
| XTTS-v2 | 2.09 GB | none published | 8 GB |
| Qwen3-TTS 12Hz 0.6B Base | 2.51 GB | none published | 8–12 GB |
| Zonos v0.1 transformer | 3.25 GB | 6 GB+ VRAM; hybrid needs 3000-series or newer | 8 GB transformer, 24 GB hybrid |
| Chatterbox (one variant) | ~3.2 GB | none published | 8 GB |
| Chatterbox-Nano | 3.00 GB | runs 3x realtime on 8 CPU cores | 4 GB or CPU |
| CosyVoice2-0.5B | 3.55 GB | none published | 8 GB |
| Qwen3-TTS 12Hz 1.7B Base | 4.54 GB | FlashAttention 2 recommended to reduce GPU memory | 12–16 GB |
| VoxCPM2 | 4.96 GB | ~8 GB | 8–12 GB |
| IndexTTS 2.5 | 5.47 GB | FP16/BF16 “lower VRAM usage” | 16 GB |
| VibeVoice 1.5B | 5.41 GB | none published | 16 GB |
| MOSS-TTSD v0.7 | 4.10 GB | none published | 12 GB |
| Fun-CosyVoice3-0.5B-2512 | 6.45 GB | none published | 16 GB |
| F5-TTS | 6.74 GB | none published | 12–16 GB |
| Parler-TTS large v1 | 9.33 GB | none published | 16 GB |
| Fish Audio S2-Pro | 10.99 GB | benchmarks on a single H200 | 24 GB+ |
| Dia 1.6B | 12.89 GB | around 10 GB of VRAM | 16 GB |
| Fish Audio Drama 3 | 41.47 GB | benchmarks on a single H200 | 80 GB+ |
The last column is my inference, not a published requirement. The rule I used: peak VRAM runs roughly 1.3 to 1.6 times on-disk weight size at fp16 or bf16, more if the pipeline loads a second model. F5-TTS is the clearest example: it needs an ASR pass to transcribe your reference clip unless you supply the text, so you hold the 6.74 GB checkpoint plus an ASR model plus a vocoder. Note the two disagreements: Dia’s weights are 12.89 GB on disk against a stated “around 10 GB of VRAM”, and Zonos recommends 6 GB against 3.25 GB of weights.
On consumer hardware, the RTX 5090 ships 32 GB of GDDR7 and the RTX 4090 remains the 24 GB workhorse, enough for everything except Fish Audio Drama 3 and other 35B-class models.
Apple Silicon is a genuinely different option
Unified memory changes the maths. Apple’s current Mac Studio ships with 48, 64 or 128 GB of unified memory on M5 Max and 256 GB or 512 GB on M5 Ultra, and the GPU shares that pool rather than owning a separate 24 GB. For a 6 GB model, a 32 GB Mac is a real deployment target.
The catch is software maturity. Kokoro documents the route: set PYTORCH_ENABLE_MPS_FALLBACK=1 on M-series hardware to get GPU acceleration. But new projects ship CUDA-first. Zonos restricts support to Linux and macOS with 3000-series-or-newer NVIDIA GPUs, with Windows only through an unofficial fork.
Windows is where voice tooling breaks
If you are on Windows and evaluating RVC or so-vits-svc, expect friction. The current RVC README targets Python 3.12 x64 and requires a deliberate two-stage install (CUDA-matched PyTorch first, then the matching requirements file) with torch==2.7.1+cu128 for RTX 50-series cards and torch==2.7.1+cu118 for older cards. Windows users also place ffmpeg.exe and ffprobe.exe in the project root, and AMD or Intel cards need DirectML. On Ubuntu 24.04 the same README installs everything through apt and moves on.
That is not a hard problem, it is a distraction from the problem you hired the model to solve.
On RunPod: pods versus serverless
RunPod’s pricing page lists two tiers per GPU. Community Cloud is the cheaper, preemptible column; Secure Cloud is the premium column for production workloads, a distinction RunPod names explicitly in its own guide to running open-source models. Rates below are from 16 October 2026.
| GPU | VRAM | Community Cloud | Secure Cloud |
|---|---|---|---|
| RTX A5000 | 24 GB | $0.27/hr | $0.69/hr |
| RTX 3090 | 24 GB | $0.50/hr | $0.69/hr |
| A40 / RTX A6000 | 48 GB | $0.59/hr | $1.22/hr |
| L4 | 24 GB | $0.59/hr | $0.69/hr |
| RTX 4090 | 24 GB | $0.89/hr | $1.10/hr |
| RTX 5090 | 32 GB | $1.19/hr | $1.58/hr |
| RTX 6000 Ada / L40S | 48 GB | $0.99–$1.09/hr | $1.75/hr |
| RTX Pro 6000 | 96 GB | $2.49/hr | $3.49/hr |
| A100 PCIe / SXM | 80 GB | $1.79/hr | $2.72/hr |
| H100 PCIe | 80 GB | $2.89/hr | $4.79/hr |
| H200 | 141 GB | $5.29/hr | $5.93/hr |
Storage adds up slowly. Container disk is $0.10/GB per month while a Pod runs; a volume disk is $0.10/GB running and $0.20/GB stopped; a network volume is $0.07/GB under 1 TB, dropping to $0.05 above that, and it persists across Pods. A stopped Pod without a network volume is terminated and its data is unrecoverable, so if you care about your checkpoints, use one.
The cold-start problem is the real design decision. RunPod’s own cold-start write-up, published 14 July 2026, states that end-to-end cold starts typically run 30 seconds to two minutes plus, that loading a 7B model from a network volume takes around 15 seconds, and that downloading the same model from Hugging Face at startup can take up to two minutes. You cannot engineer around the disk-to-VRAM part. Their FlashBoot feature keeps paused containers with weights still resident in VRAM and resumes them in about 200 ms, and a worker that has scaled down and shows as Idle is not billed at all.
If you need predictable latency, set active workers above zero. RunPod’s optimization documentation gives the sizing formula directly: active workers = (requests per minute × request duration in seconds) divided by 60. Their own example is six requests per minute at 30 seconds each needing three active workers.
The other hosts, briefly
- Lambda lists A100 SXM 80 GB at $2.79 per GPU per hour, H100 SXM 80 GB at $3.99 and B200 SXM 180 GB at $6.69, with 1-Click Clusters from $9.36 to $9.86 per GPU per hour on two-week to one-year terms.
- CoreWeave quotes whole 8-GPU nodes: HGX H100 at $49.24 per hour on-demand and $19.71 spot, roughly $6.16 and $2.46 per GPU. L40S nodes are $18.00 on-demand and $7.88 spot.
- Hugging Face lists dedicated Inference Endpoints from $0.033 per hour, with A100 80 GB at $2.50, RTX PRO 6000 96 GB at $2.75, L40S 48 GB at $1.80 and L4 24 GB at $0.80. ZeroGPU Spaces on an RTX Pro 6000 Blackwell are free.
- Modal charges per second rather than per GPU-hour: $0.00003942 per physical core per second and $0.00000667 per GiB per second, with non-preemptible execution at three times base.
- Vast.ai advertises interruptible instances at 50% or more below on-demand, per-second billing across 68+ GPU types, and reserved capacity at up to 50% off. A stable RTX 4090 spot rate is not confirmed here; check the current offers before budgeting.
- Paperspace now resolves to DigitalOcean’s Gradient and GPU Cloud pages, listing older cards such as the A4000 at $0.76 per hour and the A6000 at $1.89.
What it actually costs per finished audio hour
This is the number most comparisons skip. Because GPU rental is billed per second of wall-clock time, and real-time factor converts wall-clock time into audio duration, the formula is exact:
cost per finished audio hour = hourly GPU price × real-time factor
For the API baseline I need a characters-to-audio conversion. The Kokoro model card states that on average 1,000 characters of input is about one minute of output. That is a community estimate from a vendor card and is roughly 18 months old, so treat it as a planning assumption rather than a constant. One hour of finished audio is therefore about 60,000 characters. Against that, ElevenLabs’ API pricing lists Eleven v3 Multilingual and Eleven v4 at $0.08 per 1,000 characters, and Eleven v3 Conversational, Eleven v4 Turbo and Flash/Turbo v2.5 at $0.04 per 1,000.
| Route | List price | RTF used | Cost per finished audio hour |
|---|---|---|---|
| ElevenLabs v3 Multilingual (API) | $0.08 / 1k chars | n/a | $4.80 |
| ElevenLabs v4 (API) | $0.08 / 1k chars | n/a | $4.80 |
| ElevenLabs Flash / Turbo v2.5 (API) | $0.04 / 1k chars | n/a | $2.40 |
| RunPod 4090 Community + VoxCPM2 + Nano-VLLM | $0.89/hr | 0.13 | $0.12 |
| RunPod 4090 Community + IndexTTS 2.5 (bf16) | $0.89/hr | 0.2065 | $0.18 |
| RunPod 4090 Community + VoxCPM2 (standard) | $0.89/hr | ~0.30 | $0.27 |
| RunPod 4090 Secure + VoxCPM2 (standard) | $1.10/hr | ~0.30 | $0.33 |
| RunPod 4090 Community + Zonos v0.1 | $0.89/hr | ~0.50 | $0.45 |
| RunPod 4090 Community + Dia 1.6B | $0.89/hr | ~2.15 | $1.91 |
Where those RTFs come from, and their limits:
- VoxCPM2 publishes RTF of about 0.30 on an RTX 4090 and about 0.13 with Nano-VLLM, plus a stated VRAM requirement of about 8 GB.
- IndexTTS 2.5 publishes a full RTF table measured on an RTX 4090 with KV cache enabled: 0.2065 overall in bf16 and 0.2060 in fp32, against 0.3257 and 0.3748 for the earlier IndexTTS-2 in fp16 and fp32.
- Zonos describes itself as running “with a real-time factor of ~2x on an RTX 4090”, which would imply RTF ≈ 0.5 if it means twice real-time speed. The wording is ambiguous, without a benchmark table.
- Dia is the outlier. Its card states that on an A4000 it generates roughly 40 tokens per second where 86 tokens equals one second of audio, an RTF of about 2.15: slower than real time.
- F5-TTS publishes RTF 0.0394 at concurrency 2 and 253 ms average latency on a single L20 through Triton and TensorRT-LLM, and 0.1467 offline in plain PyTorch. That is an L20, not a 4090.
Three caveats matter more than the arithmetic. These are vendor measurements on vendor hardware, not mine. The figures assume a saturated GPU (no idle time, no cold starts, no failed jobs) so real batch throughput will be worse. And the per-character conversion varies with speaking rate and language.
The break-even question
A $0.89 per hour Community Cloud 4090 running every hour of the month costs about $650. At an RTF of 0.30 that buys roughly 2,430 finished hours of audio a month, or about 146 million characters. Below that volume, renting beats buying.
How to deploy a model on RunPod
-
Pick the licence first, not the model. Read the
LICENSE.txtin the Hub repo, not the GitHub repo. If the weights are NC, NC-SA, a research licence or a custom agreement with a revenue ceiling, stop. -
Match the GPU to the weights, with headroom. Read the weight size off the Hub, double it for activations, vocoder and any second model the pipeline loads, and buy the next card up. A 24 GB RTX 4090 covers everything in the table except Fish Audio Drama 3 and other 35B-class models.
-
Deploy a Pod, not Serverless, for the first run. Pick Community Cloud for the cheapest rate, attach a network volume so checkpoints survive a stop, and remember that a stopped Pod without one is terminated with unrecoverable data. RunPod charges per second and has no data ingress or egress fees, so exploratory work is cheap.
-
Use a container image, not a manual install. F5-TTS and IndexTTS both publish Dockerfiles, and CosyVoice ships gRPC and FastAPI server images. This is the biggest lever on the Windows pain described above, because the container is Linux and the dependency chain is already solved.
-
Cache the model, and know which mode you chose. RunPod ranks cached models on Hugging Face as the fastest cold start, baking models into the image as fastest for private repos, and network volumes as most flexible and worst for cold start: reading from a network volume happens after the worker starts and after billing begins.
-
Measure RTF on your own text before you commit to a host. Generate two minutes of your actual script, measure the real-time factor, then multiply by the hourly price. Vendor RTFs vary with text length, as IndexTTS’s own table shows.
-
Move to Serverless only when traffic is steady. If request volume justifies it, set active workers above zero using RunPod’s formula and turn on FlashBoot. If traffic is spiky, leave flex workers to scale to zero and accept the cold start as the price of not paying for idle GPUs.
-
Decide on watermarking and disclosure before launch, not after. Chatterbox embeds a Perceptual Threshold watermark in every generated file by design, and the Fish Audio licence prohibits removing or obscuring provenance signals. If your output reaches EU users, the Article 50 labelling duty has applied since 2 August 2026.
What breaks: failure modes and gotchas
The licence assumption. The most common failure is reading a code licence and assuming it governs the weights. F5-TTS, Spark-TTS and XTTS-v2 all have this gap, and in F5-TTS’s case the code is MIT and the weights are not.
Quantised community checkpoints. Quantised variants exist on the Hub for most of these models and they are not reviewed by the maintainers. Evaluate voice similarity on the full-precision weights.
Reference audio quality dominating everything. Nearly every zero-shot model here improves dramatically with a clean 10 to 30 second reference. Chatterbox warns that if the reference clip’s language does not match the target language tag, the output can inherit the reference’s accent unless you set the CFG weight to 0. IndexTTS 2.5 lets you pass the reference transcript for higher fidelity.
Serving stack, not model, sets the bill. VoxCPM2 reports RTF of about 0.13 with Nano-VLLM against about 0.30 standard, a 2.3x difference that comes entirely from the runtime. Qwen3-TTS recommends FlashAttention 2 specifically to reduce GPU memory usage. Model card numbers and production numbers are different numbers.
Batching trades latency for throughput. Fish Audio Drama 3 publishes scaling figures: on one H200, mean RTF is 0.131 at concurrency 1 and 0.801 at concurrency 64.
The unfrozen project. The Coqui TTS repository has 46,000 stars and its last push was 16 August 2024. XTTS-v2 will keep working, and you will not get security fixes, Python 3.13 support or CUDA 12.8 wheels for it.
Serverless surprise bills. RunPod’s default account spend limit is $80 per hour and Pods stop automatically when your balance hits $0. Set the limit deliberately rather than discovering it.
Voice cloning law: consent, publicity rights and Article 50
This section summarises published rules. Have the consent and licensing terms reviewed for your intended use.
Tennessee’s ELVIS Act (the Ensuring Likeness Voice and Image Security Act of 2024) was passed on 21 March 2024 and took effect on 1 July 2024, and a Vanderbilt Law Review article notes it was the first AI-focused US state law. Reed Smith describes it as granting rights over an individual’s voice “in any medium and in any manner, regardless of whether the sound contains the actual voice or a simulation”, and extending liability beyond end users to technology providers whose primary purpose is producing a particular identifiable individual’s photograph, voice or likeness without authorisation. Resemble AI’s reference index adds treble damages for knowing violations and coverage of deceased individuals’ estates.
California’s AB 1836 was signed on 17 September 2024 and took effect on 1 January 2025. According to Manatt, it amends Civil Code § 3344.1 to create a cause of action against anyone who produces, distributes or makes available a digital replica of a deceased personality’s voice or likeness without consent, with damages of the greater of $10,000 or actual damages. Exemptions cover news reporting, comment, criticism and parody, fleeting or incidental use, and historical or biographical use that does not falsely imply an authentic recording. The companion bill AB 2602 makes contract clauses unenforceable in defined circumstances where a performer had no legal or union representation.
The federal TAKE IT DOWN Act was signed on 19 May 2025. The FTC began enforcing Section 3 on 19 May 2026: covered platforms must remove nonconsensual intimate imagery, including AI-generated “digital forgeries”, and known identical copies within 48 hours of a valid request, with potential civil penalties of $53,088 per violation. It targets platforms rather than model runners, but it shapes platform terms of service.
The EU AI Act Article 50 has applied since 2 August 2026, according to the European Commission’s own FAQ page, last updated 24 July 2026. Four duties matter for speech:
- Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format so they are detectable as artificially generated or manipulated. Systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with this particular duty.
- Article 50(4) requires deployers to disclose that content constituting a deepfake has been artificially generated or manipulated, at the latest at first exposure, in a clear and distinguishable form. The Commission’s guidance is explicit that you cannot rely only on the Article 50(2) machine-readable marking: you need a perceptible audible or visible label.
- Article 3(60) defines a deepfake by three cumulative criteria: resemblance to the subject, the subject existing or being capable of existing, and the content falsely appearing authentic or truthful. A synthetic voice nobody could mistake for a real person is arguably outside that definition.
- Penalties reach up to EUR 15 million or 3% of total worldwide turnover, with proportionality considered for SMEs.
A voluntary Code of Practice on Transparency of AI-generated content has been assessed as adequate by both the Commission and the AI Board, and signatories get legal certainty regardless of establishment or operating location.
- Penalties reach up to EUR 15 million or 3% of total worldwide turnover for the preceding financial year, with proportionality considered for SMEs.
A voluntary Code of Practice on Transparency of AI-generated content has been assessed as adequate by both the Commission and the AI Board, and signatories get legal certainty regardless of establishment or operating location. Providers who do not sign it must demonstrate compliance by other means and may face more information requests.
The practical read: if you generate a recognisable clone of a real, identifiable person and distribute it in the EU, you owe a visible disclosure. A designed voice that resembles nobody is in a different and much safer place.
The bottom line
Start with the licence, because it is free and it is the only constraint you cannot engineer around. If you need paid output today, Kokoro-82M, Chatterbox, VoxCPM2, Qwen3-TTS, CosyVoice2 and Fun-CosyVoice3 are all unambiguous.
Then size by VRAM, which is measurable: read the weight size off the Hub, double it, and buy the next card up. A 24 GB card covers essentially the whole commercially usable landscape.
Then buy the economics. At $0.89 per hour, a RunPod Community Cloud RTX 4090 produces roughly $0.27 per finished audio hour at RTF 0.30: an order of magnitude under ElevenLabs list pricing, and it keeps going down as serving stacks improve.
And if you are picking between a Linux Pod and your own Windows machine, pick the Pod. The dependency chain is not the interesting part of your job.
FAQ
Can I use XTTS v2 commercially?
No, not without a separate licence you can no longer buy. The XTTS-v2 weights are shipped under the Coqui Public Model License 1.0.0, whose first line reads that it allows only non-commercial use of a machine learning model and its outputs. Coqui AI, the company that published it, shut down in early 2024, so there is no functioning commercial licensing channel. For paid work, pick an Apache-2.0 or MIT model such as Kokoro-82M, Chatterbox, VoxCPM2 or CosyVoice2 instead.
Which open-weight TTS models are safe for commercial use?
The clearest yes-list is Kokoro-82M, Chatterbox, VoxCPM2, Qwen3-TTS, CosyVoice2 and Fun-CosyVoice3, all Apache-2.0 or MIT, plus OpenVoice V2 under MIT. Fish Audio Drama 3 is conditional: the Community License v1.0 dated 13 October 2026 permits commercial use only under USD 500,000 of annual revenue and forbids offering it as a hosted service. IndexTTS 2.5 uses the bilibili Model Use License Agreement, commercial below 100 million MAU and RMB 1 billion in revenue.
How much VRAM do I need to run these models?
Weight size on the Hub is a useful starting point, and most pipelines fit in 1.3 to 1.6 times the on-disk weight size at fp16 or bf16. Kokoro-82M at 0.36 GB runs on any 4 GB card, XTTS-v2 at 2.09 GB fits in 8 GB, and F5-TTS at 6.74 GB wants 12 to 16 GB because it also loads a vocoder and an ASR model. Three projects publish their own numbers: Zonos 6 GB or more, VoxCPM2 about 8 GB, Dia 1.6B around 10 GB.
Is RunPod actually cheaper than ElevenLabs?
For high volume, yes, by roughly an order of magnitude. ElevenLabs lists Eleven v3 Multilingual at $0.08 per 1,000 characters and Flash v2.5 at $0.04. Using about 1,000 characters per minute of audio, one finished hour costs roughly $4.80 on v3 Multilingual and $2.40 on Flash v2.5. A RunPod Community Cloud RTX 4090 at $0.89 per hour at a real-time factor of 0.30 works out to about $0.27 per finished audio hour.
What is a real-time factor and why does it decide my bill?
Real-time factor, or RTF, is the wall-clock time to generate one second of audio divided by the length of that audio. An RTF of 0.30 means three seconds of audio per second of compute. Since GPU rental is billed per second of wall-clock time, your cost per finished audio hour equals the hourly GPU price multiplied by the RTF.
Should I use RunPod Pods or RunPod Serverless?
Use a persistent Pod for development, batch work and anything spiky. Use Serverless for steady low-latency traffic. RunPod states that serverless workers are billed from the moment they start, including the cold start, and that end-to-end cold starts typically run 30 seconds to two minutes plus. Flex workers scale to zero for free but pay that cold start again after a quiet period; active workers run 24/7 and eliminate it at the price of always-on GPU spend.
Do I need permission to clone someone’s voice?
Yes, and in several jurisdictions you need more than permission. Tennessee’s ELVIS Act, effective 1 July 2024, was the first US state law to explicitly cover AI voice cloning, extending the state’s right of publicity to voice for any individual rather than performers alone. California’s AB 1836, effective 1 January 2025, creates a cause of action over digital replicas of deceased performers with damages of the greater of $10,000 or actual loss.
Does the EU AI Act require me to label AI-generated voice?
For deepfakes, yes, and it is already live. Article 50 of the EU AI Act has applied since 2 August 2026. Deployers must disclose that audio, image or video constituting a deepfake has been artificially generated or manipulated, at the latest at first exposure, in a clear and distinguishable form. You cannot satisfy this with a machine-readable watermark alone. Penalties reach up to EUR 15 million or 3% of worldwide turnover.
Can I run these models on a Mac instead of a cloud GPU?
Yes for the smaller models, with caveats. Apple Silicon shares one memory pool, so a 36 GB Mac Studio can allocate most of it to a model where a discrete card cannot. Kokoro tells Mac users to set PYTORCH_ENABLE_MPS_FALLBACK=1 to get GPU acceleration on M-series hardware. Expect MPS to be slower than CUDA, and expect new projects to ship CUDA-first instructions, as Zonos does.
Why does Windows make voice conversion so much harder?
Because voice-conversion tooling depends on compiled binaries that mostly ship Linux wheels. The current RVC README targets Python 3.12 x64 and needs a two-stage install: CUDA-matched PyTorch first, then the matching requirements file, with CUDA 12.8 for RTX 50-series cards and CUDA 11.8 for older cards. Windows users also drop ffmpeg.exe and ffprobe.exe into the project root.
Sources
- Coqui Public Model License 1.0.0 (XTTS-v2 LICENSE.txt): the non-commercial clause governing XTTS-v2 weights and outputs
- coqui/XTTS-v2 model card: 17 languages, 6-second cloning, links to CPML
- coqui-ai/TTS GitHub repository: MPL-2.0 code licence, Windows install note, Python version support
- SWivid/F5-TTS GitHub repository: code under MIT, pre-trained models under CC-BY-NC, L20 real-time factor benchmarks
- F5 Audio Community License Agreement v1.0: USD 500,000 revenue threshold, Hosted Service prohibition, last updated 13 October 2026
- fishaudio/s2-pro model card: Fish Audio Research Licence, H200 RTF 0.195, 80+ languages
- fishaudio/drama-3 model card: 35B MoE, concurrency table on one H200, Community Licence
- FunAudioLLM/CosyVoice2-0.5B model card: Apache-2.0, evaluation tables against F5-TTS and Spark TTS, streaming latency claim
- QwenAudio/CosyVoice GitHub repository: Apache-2.0 code licence, vLLM and Docker deployment paths
- Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice model card: 10 languages, FlashAttention 2 memory advice, benchmark tables
- ResembleAI/chatterbox GitHub repository: MIT, Turbo and Nano sizes, 23 languages, PerTh watermarking
- openbmb/VoxCPM2 model card: Apache-2.0, about 8 GB VRAM, RTF 0.30 and 0.13 on RTX 4090, 30 languages
- index-tts/index-tts GitHub repository: RTX 4090 real-time factor tables, release dates, bilibili licence pointer
- bilibili Model Use License Agreement: 100M MAU and RMB 1 billion revenue thresholds, prohibited high-risk uses
- Zyphra/Zonos GitHub repository: Apache-2.0, 6 GB+ VRAM, 2x realtime on RTX 4090, platform support
- hexgrad/Kokoro-82M model card: Apache-2.0, 82M parameters, characters-per-minute conversion and third-party API rate comparison
- hexgrad/kokoro GitHub repository: Windows espeak-ng install, MPS fallback guidance
- nari-labs/Dia-1.6B model card: Apache-2.0, around 10 GB VRAM, A4000 token rate
- boson-ai/Higgs-Audio GitHub repository: Boson Higgs Audio v3 Research and Non-Commercial License
- 2noise/ChatTTS GitHub repository: AGPL-3.0 code licence and academic-use-only model disclaimer
- OHF-Voice/piper1-gpl: GPL-3.0 successor engine, used by Home Assistant and NVDA
- rhasspy/piper-voices on Hugging Face: voice file repository, 54 language directories
- myshell-ai/OpenVoiceV2 model card: MIT from April 2024, commercial use permitted
- yl4579/StyleTTS2 GitHub repository: MIT code, disclosure-and-consent terms on pretrained models
- RVC-Project/Retrieval-based-Voice-Conversion-WebUI: MIT, Python 3.12 x64, two-stage CUDA 12.8 and 11.8 install, ffmpeg on Windows
- RunPod GPU pricing: Community Cloud and Secure Cloud hourly rates by GPU and VRAM
- RunPod Pods pricing documentation: storage rates, stop versus terminate behaviour, $80 per hour default spend limit
- RunPod Serverless pricing documentation: per-second billing, three billed phases, 5-second default idle timeout
- RunPod Serverless optimization guide: active worker formula, model caching hierarchy, last modified 15 October 2026
- RunPod Serverless workers overview: worker lifecycle, active versus flex, max worker limits by account balance
- Cold starts were never the real problem (RunPod blog, 14 July 2026): 30 seconds to 2 minutes cold starts, FlashBoot 200 ms resume
- RunPod: Try Open-Source AI Models Without Installing Anything Locally: Community Cloud versus Secure Cloud definitions
- Lambda GPU cloud pricing: A100, H100 and B200 per GPU per hour, 1-Click Cluster tiers
- CoreWeave cloud pricing: 8-GPU node on-demand and spot prices for H100, L40S, A100
- Hugging Face pricing: Inference Endpoints from $0.033 per hour, GPU instance rates
- Modal pricing: per-second CPU and memory rates, Starter and Team plans, non-preemptible multiplier
- Vast.ai GPU pricing: interruptible at 50% or more below on-demand, per-second billing
- ElevenLabs API pricing: per 1,000 character rates for v4, v3 and Flash/Turbo v2.5
- DigitalOcean (Paperspace) GPU Cloud pricing: A4000, A6000 and H100 hourly rates, Gradient plans
- NVIDIA GeForce RTX 5090 specifications: 32 GB GDDR7, 575 W board power
- Apple Mac Studio technical specifications: unified memory configurations up to 512 GB
- EU AI Act Article 50, AI Act Service Desk: full statutory text of paragraphs 1 to 7
- European Commission FAQ on Article 50 transparency obligations: 2 August 2026 application date, 2 December 2026 grace period, EUR 15 million or 3% penalties, last updated 24 July 2026
- European Commission guidelines on Article 50: official guidelines page and scope definitions
- FTC: Take It Down Act enforcement starts now: 19 May 2025 signature, 19 May 2026 enforcement, 48-hour removal, $53,088 per violation
- Why Tennessee’s ELVIS Act Is the King of AI Protections (27 Vand. J. Ent. & Tech. L. 259): passage date 21 March 2024 and effective date 1 July 2024
- Reed Smith: Tennessee first in efforts to protect ROP against GenAI imitations: scope of liability for developers and providers of cloning technology
- Resemble AI deepfake law reference: Tennessee ELVIS Act: treble damages, coverage of deceased individuals’ estates, companion statutes
- Manatt: California Enacts a Suite of New AI and Digital Replica Laws: AB 1836 and AB 2602 details, 17 September 2024 signature, 1 January 2025 effect
