In August 2026, Resemble AI retired Chatterbox, an MIT-licensed text to speech model with 27.9k stars on GitHub, and Resemble AI says the code stays public "with no support, updates, or paid tier." Free to download turned out not to mean free to keep.
Open-source text to speech means a team downloads a model's weights, the learned numbers that let it speak, and runs them on its own servers. A hosted API means a company runs the model, and a developer sends text over the internet and pays for what is generated. Kokoro, Orpheus, Qwen3-TTS, and Chatterbox are downloadable models. Bland Speech, ElevenLabs, Cartesia, and Amazon Polly are hosted.
What is the difference between open-source text to speech and a hosted API?
Open-source text to speech such as Kokoro, Orpheus, and Qwen3-TTS gives a team model weights to run on its own hardware, while a hosted API such as Bland Speech, ElevenLabs, or Cartesia runs the model and bills for use, as of November 2026.
The weights carry no per-character fee, so the bill turns into servers, engineers, and on-call hours. Model size shows what that means in hardware. Kokoro has 82 million parameters, Chatterbox Nano has 110 million and runs, in Resemble AI's words, "3x faster than realtime on 8 CPU cores", and Orpheus is built on a 3-billion-parameter Llama backbone with a realtime example that runs on vLLM, a model-serving engine.
A hosted API turns that into a request. A developer calls Bland's POST /v2/tts endpoint for a finished clip or WSS /v2/tts/ws for streaming, and pays $0.04 per 1,000 characters, pay as you go with no monthly plan, with 50,000 free characters once for a new account. Bland's text to speech API page holds the endpoint details.
A third route sits between the two, which is a downloadable model that someone else hosts. Kokoro's author wrote that "as of April 2025, the market rate of Kokoro served over API is under $1 per million characters of text input", and the same author warns that any website with "kokoro" in its domain is not affiliated with the model. Canopy Labs calls Baseten its "preferred inference partner" for Orpheus.
Can open-source text to speech be used commercially?
The answer depends on the licence: Kokoro, Orpheus, and Qwen3-TTS ship under Apache 2.0, Chatterbox under MIT, and XTTS-v2 under terms that allow only non-commercial use.
What these rows mean
- Licence: the name each maker gives the licence on its model card, repository, or licence file.
- What it allows: quoted or paraphrased from that text. A commercial team reads the full licence text before building on any of these models.
- Fish Audio: hosted services and commercial use above the revenue cap need a separate licence from Fish Audio.
The Coqui row carries the sharpest catch. The company shut down in January 2024, the XTTS-v2 licence allows "only non-commercial use of a machine learning model and its outputs", and the licence text points to a page on Coqui's domain that now returns an error.
What happens to open-source text to speech when its maintainer stops?
The code keeps running and nothing else moves: Chatterbox was retired in August 2026, Coqui shut down in January 2024, Kokoro's last GitHub commit is from August 2025, and Orpheus's is from December 2025.
Canopy Labs now describes itself as building "realtime interactive models for avatars" and says its next speech models are "currently, but not for long, private". Canopy Labs still has a cloning notebook to fix and a realtime streaming glitch that "occasionally skips frames" among its open items. For XTTS, Idiap Research Institute maintains a community fork of the Coqui code and calls the original "unmaintained", with a release as recent as 19 October 2026, but a fork of the code does not change what the model licence says.
Hosted APIs end too, and recently. LMNT says "LMNT has shut down". OpenAI says its tts-1, tts-1-hd, and gpt-4o-mini-tts models shut down on 6 January 2027. Resemble AI says it is "not selling voice AI to new customers". The difference is who carries the problem: a vendor notice puts a migration on the team's calendar, and an abandoned open model leaves a team to patch it. The provider shutdowns explainer covers each of those endings and how a team moves.
What does a team have to build around an open-source model?
A team builds the serving layer itself: Kokoro's example is a Python generator that returns audio chunks, Chatterbox's is a model.generate(text) call that returns a full waveform, and Kokoro, Orpheus, and XTTS-v2 all write 24 kHz audio in their examples.
Streaming varies by model, and a phone line varies again. Speechify says Twilio and LiveKit SIP expect 8 kHz mu-law or 16 kHz PCM, so 24 kHz output is audio a team converts before it reaches a call. Qwen3-TTS says it can output its first audio packet after a single character and quotes end-to-end synthesis latency "as low as 97ms", and Coqui said XTTS "can now stream with <200ms latency". Orpheus quotes "~200ms streaming latency", reducible to about 100ms with input streaming. Each of those is a claim from the maker, and a team's own GPU decides what it hears.
Two smaller things come with running weights. Kokoro's author notes "weakness on short utterances, especially less than 10-20 tokens", while Design Arena notes that phone-agent prompts are single turns that "tend to be shorter". Self-run weights carry no vendor SOC 2, HIPAA, or retention terms, because the team's own infrastructure is the whole stack.
A hosted API ships that layer finished. Bland Speech streams over HTTP and a WebSocket, accepts text as a language model writes it, lets a new context_id interrupt the turn that is playing, returns 8 kHz mu-law for phone lines, and answers on an OpenAI-compatible endpoint. Pipecat maintains a Bland service and LiveKit offers a Python plugin. Bland is SOC 2 Type II audited, and Bland says speech synthesis runs inside its FedRAMP boundary.
Which should a team choose, open-source text to speech or a hosted API?
Open-source text to speech suits a team that must keep audio on its own hardware and has engineers to run it, while a hosted API such as Bland Speech suits a team shipping a phone voice agent that needs streaming, interruption, and phone audio working on day one.
Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and the human recording scored 1500. None of those open models was in those results. The price is higher per character than the alternatives: Amazon Polly Standard is $4.00 per million characters, Azure Neural is $15 per million, and Bland Speech is $40 per million.
Use Bland Speech when
- A team is shipping a phone voice agent and will call one API for streaming, interruption, and mu-law audio
- A developer is moving a voice agent off a retired or deprecated text to speech model and wants a hosted replacement
- A buyer will pay per character and keep servers, model updates, and on-call out of scope
Use open-source text to speech when
- A team will run GPUs, patch the model, and write its own streaming server
- Audio must stay on hardware the team controls, and the licence fits the product
- An engineer is building a prototype or research project where a maintained service is not required
Pros
- One API for HTTP and WebSocket streaming, interruption, and mu-law phone audio
- A model that ranked first of 13 AI voices in Design Arena's results of 4 August 2026
- Per-character billing at $0.04 per 1,000 characters, with 50,000 free characters for a new account
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
The wider field is in the open-source text to speech comparison, the Kokoro comparison and the Chatterbox comparison go model by model, and the text to speech hub holds the rest. More explainers sit under text to speech guides, and Bland Speech has a studio to hear the voice.
Frequently asked questions
Is open-source text to speech free for commercial use?
Open-source models are free to download, and commercial use depends on the licence. Kokoro and Qwen3-TTS use Apache 2.0, Chatterbox uses MIT, Fish Audio's community licence covers commercial use only while annual revenue is under USD 500,000, and XTTS-v2 allows only non-commercial use. Running the model still costs servers and engineers.
Does Kokoro have an official hosted API?
No. Kokoro is trained by an individual author, who warns that any website with "kokoro" in its root domain is not affiliated with the model, so a hosted Kokoro comes from a third party. In April 2025 the author put the market rate for Kokoro over an API under $1 per million characters.
What happened to Coqui and XTTS?
Coqui shut down in January 2024, and its old Studio and API are gone. The XTTS-v2 model remains on Hugging Face under a non-commercial licence, and Idiap Research Institute maintains a community fork of the code.
Is Chatterbox still maintained?
No. Resemble AI says "As of August 2026, Chatterbox is retired research and is no longer maintained." The code stays public under the MIT licence. Resemble AI's Hugging Face card still points to a paid service that Resemble AI says does not exist.
Is Bland Speech open source?
No. Bland operates a closed-source model it developed itself, and a team reaches Bland Speech v3 through /v2/tts, /v2/tts/ws, and an OpenAI-compatible endpoint. A team that needs weights on its own hardware picks one of the open models above.
Can a hosted text to speech API run on dedicated hardware?
For Bland Speech, Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland. Enterprise pricing is custom, set with Bland's sales team.
Which costs less per character?
Hosted open models and standard cloud voices cost less than Bland Speech per character. Amazon Polly Standard is $4.00 per million characters and Azure Neural is $15 per million, against $40 per million for Bland Speech, and a downloaded model has no per-character fee but needs servers. Bland Speech v3 ranked first of 13 AI voices in Design Arena's results of 4 August 2026, which is what the higher rate buys.
