Canopy Labs says Orpheus TTS streams with "~200ms streaming latency for realtime applications", and names no hardware beside the figure. For a model a team hosts itself, that matters more than the number does, because latency is a property of the machine under the model as much as the model.
Bland Speech is the opposite arrangement: a text to speech API that Bland runs, with a published figure of 199ms to first audio and a price per character. The choice between the two is a choice between buying a result and owning a project.
Which is better, Bland Speech or Orpheus TTS?
For production phone calls, Bland Speech is the better choice as of November 2026. Bland runs the model, prices it at $0.04 per 1,000 characters, and returns phone-ready audio, while Orpheus TTS is free weights that a team hosts and tunes itself. Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, and Orpheus TTS wasn't in those results.
Orpheus TTS makes sense for a team that wants to fine-tune a speech model on its own recordings under the Apache 2.0 licence and has engineers to run it. A team weighing several open models can start with open-source text to speech compared.
Use Bland Speech when
- Callers on a phone line should hear a voice that sounds like a person
- A team wants one standard per-character rate and no GPUs to size, patch, or keep warm
- An existing OpenAI audio speech integration should move by changing a base URL, a key, and a model name
- A script needs performed delivery, such as a laugh or a warm read, written inline as tags
- A team wants a private voice clone from one 10-second sample, with consent confirmed for every voice
Use Orpheus TTS when
- A team will fine-tune a voice on its own data, starting from about 50 examples and aiming for 300 per speaker as Canopy recommends
- A team will run a 3B-parameter Llama-based speech model on its own GPUs, or on CPU through llama.cpp
- A team will build its own streaming server around the vLLM example in Canopy's repository
- A team will carry its own upkeep, because the repository has had no commit since 5 December 2025
What these rows mean
- Where it runs: Bland Speech is called over HTTP or a WebSocket. Orpheus TTS is weights on Hugging Face under
canopylabs, and Canopy names Baseten as its "preferred inference partner" for hosted deployments. - Audio Realism Bench, 4 August 2026: Design Arena's blind listening test, where listeners hear two clips of the same line from two models and pick the one that sounds more like a real person. A higher score means listeners picked that voice more often.
- Price: Bland Speech costs $0.04 per 1,000 characters in the studio or the API. Orpheus weights are free, so a team pays for its own machine, or for a partner's hosting.
- First audio: Each figure is the company's own published claim. Canopy's line is "~200ms streaming latency for realtime applications, reducible to ~100ms with input streaming", and Bland publishes 199ms to first audio on a streamed request.
- Phone audio: Mu-law is an 8-bit telephony format. Canopy's example code writes a WAV file at 24,000 Hz, mono, 16-bit.
- Voice cloning: Canopy says to put text and speech pairs in the prompt of the pretrained model, and "Fix voice cloning Colab notebook implementation" is still open on Canopy's checklist.
How do Bland Speech and Orpheus TTS compare?
Who carries the model is what separates these two. Bland runs Bland Speech and answers for it, while Canopy released Orpheus TTS in March 2025 and now describes itself as a company "building realtime interactive models for avatars".
Bland Speech overview
Bland Speech is the text to speech model inside Bland, a voice AI provider that runs its own model and speech rather than wrapping third parties, and Bland Speech is a closed-source model Bland developed itself. It is sold on its own as a studio and an API on the Bland Speech page.
A team sends text over HTTP to POST /v2/tts, or streams language model output into WSS /v2/tts/ws, where a new context interrupts the turn that is playing. POST /v2/audio/speech follows OpenAI's audio speech contract, and the text to speech API page has the code. Tags in square brackets, such as [laughs], [sighs], or [say warmly], tell the model how to deliver the words after them, and two controls from 0.0 to 1.0, expressiveness and stability, set how much intonation and rendering vary.
Pros
- Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026
- One standard rate, $0.04 per 1,000 characters, with no monthly plan
- Bland runs the model, so a team has no inference stack to build
- An OpenAI-compatible endpoint, so an existing integration moves by changing a base URL
- Mu-law at 8 kHz and PCM from 8 to 48 kHz, ready for phone calls
- A first-party Pipecat service and a Python plugin for LiveKit Agents
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
Orpheus TTS overview
Orpheus TTS is an open-source text to speech system built on a Llama 3B backbone, with the finetuned model on Hugging Face as orpheus-3b-0.1-ft and Llama-3.2-3B-Instruct listed as its base model. Canopy ships two English models, a finetuned one for everyday use and a pretrained base trained on 100,000-plus hours of English speech, plus a multilingual research preview covering Chinese, Hindi, Korean, Spanish, Italian, French, and German.
Canopy's site says "Last year we released Orpheus TTS", and that its newer speech and avatar models are currently private. The repository has 6.4k stars and an Apache-2.0 licence, and its last commit is dated 5 December 2025.
Pros
- Apache-2.0 weights, free to download
- Tags in the text prompt, such as <laugh>, <sigh>, and <gasp>, plus eight named English voices
- A streaming example in the repository, built on vLLM
- A llama.cpp path for running without a GPU
- Fine-tuning notebooks, with Canopy saying about 50 examples start to show quality
- Optional watermarking through Silent Cipher
Cons
- Hosted deployments go through a partner, Baseten, which Canopy calls its preferred inference partner
- A streaming glitch that occasionally skips frames is still open on Canopy's checklist
- A cloning notebook fix is still open on the same checklist
- No tagged releases, and no commit since 5 December 2025
What does running Orpheus TTS take?
Running Orpheus TTS takes a GPU server, a vLLM install, and someone who watches the versions. Canopy's streaming example installs orpheus-speech, which "uses vllm under the hood for fast inference", and Canopy warns that a vLLM release was "slightly buggy", with the workaround of reverting to vllm==0.7.3.
Canopy names Baseten for a "dedicated deployment with real-time streaming", at fp8 for speed or fp16 for full fidelity, and points to a llama.cpp path for machines without a GPU. A team learns what GPU it needs, and how many calls that GPU holds at once, by testing.
Bland Speech leaves no hardware with the buyer. Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland.
How do Bland Speech and Orpheus TTS handle streaming, latency, and phone audio?
Both stream, and the difference is who built the pipe. Bland Speech streams raw audio over HTTP or a WebSocket that accepts text as a language model writes it, and Bland publishes 199ms to first audio. Orpheus's repository includes a realtime example that yields audio chunks, and Canopy claims about 200ms, reducible to about 100ms when text streams in as well.
Bland Speech returns mu-law at 8 kHz for phone lines, or PCM at 8, 16, 24, 44.1, or 48 kHz. Canopy's example code saves 24 kHz mono audio, so a team that wants a phone carrier's format adds that step itself.
How do Bland Speech and Orpheus TTS handle voice cloning and security?
Bland Speech clones from one clean sample of about 10 seconds, and a professional clone trains on 30 minutes or more of audio and passes Bland's consent and rights review. Every clone is private to the customer's organization, and the customer confirms the voice is theirs to use.
Orpheus clones by placing text and speech pairs in the prompt, and Canopy says "the more text-speech pairs you pass in the prompt, the more reliably it will generate in the correct voice". Canopy asks users not to use the models "for impersonation without consent, misinformation or deception", and watermarking is an add-on a team switches on.
On security, Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted 17 November 2026, and Bland's trust page says speech synthesis runs inside the FedRAMP boundary. Orpheus TTS comes with no vendor compliance terms, because the weights run on a team's own infrastructure and that team carries the audit.
How much do Bland Speech and Orpheus TTS cost?
Orpheus TTS is free to download and paid for in GPU hours, while Bland Speech costs $40 per million characters with nothing to host. The weights are Apache-2.0, and hosted deployments through a partner such as Baseten bill at the partner's prices.
Bland Speech costs $0.04 per 1,000 characters, in the studio or the API, with no monthly plan. New accounts get 50,000 characters free, and credits don't expire. A million characters costs $40, ten million cost $400, and a 10-minute narration costs about $0.36.
Enterprise pricing is custom, set with Bland's sales team.
What a team pays for Orpheus is a GPU, or a partner's hosting, plus the engineering time that fine-tuning, streaming, and version pinning take. Whether $40 per million characters beats that bill depends on how busy the GPU stays, and only a team's own load test answers that.
Which sounds more human, Bland Speech or Orpheus TTS?
Design Arena's Audio Realism Bench results of 4 August 2026 scored Bland Speech and not Orpheus TTS. In those results Bland Speech v3 ranked first of 13 AI voices at 1365, and real human recordings scored 1500.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 4thOpenAIGPT Realtime 21119
- 5thGoogleGemini 2.5 Pro TTS1109
- Not rankedOrpheus TTS's own voicesNot ranked
Ranks are among 13 AI voices, and Orpheus TTS's own voices weren't ranked. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
Canopy says Orpheus has speech "superior to SOTA closed source models", which is Canopy's own claim. Listeners in the Audio Realism Bench results Design Arena published hear two clips of the same line without knowing which model made which, and pick the one that sounds more like a real person.
Why is Bland Speech the best Orpheus TTS alternative?
A team that wants a voice on a phone line is buying a result, and Bland Speech sells the result: one endpoint, one standard rate, phone-ready audio, and a voice that ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026.
A team already on an OpenAI-style speech client moves by changing a base URL, a key, and a model name, and a team that fine-tuned an Orpheus voice re-creates it as a Bland clone from the original recordings, because a cloned voice isn't portable between providers. Comparing other open models is easier with Bland's Bland Speech vs Kokoro page, and the wider field sits in Bland's comparison of text to speech APIs and its head-to-heads.
Frequently asked questions
Is Bland Speech better than Orpheus TTS?
For phone calls and voice agents, yes. Bland Speech is a hosted API with a price, phone-ready audio, and a result on Audio Realism Bench, where Bland Speech v3 ranked first of 13 AI voices on 4 August 2026. Orpheus TTS wasn't in those results.
Is Orpheus TTS free?
The weights are free under Apache 2.0, and running them costs a team a machine and engineering time. Bland Speech costs $0.04 per 1,000 characters, which is $40 per million, and new accounts get 50,000 characters free.
How fast is Orpheus TTS?
Canopy claims about 200ms of streaming latency, reducible to about 100ms with input streaming. Bland publishes 199ms to first audio for Bland Speech. Each figure is the company's own claim.
Does Orpheus TTS work on phone calls?
Canopy's example code saves 24 kHz mono audio, and a team handles conversion for a phone line. Bland Speech returns mu-law at 8 kHz as well as PCM from 8 to 48 kHz, so audio goes onto a phone line without conversion.
Can Orpheus TTS clone a voice?
Canopy claims zero-shot cloning, done by putting text and speech pairs in the prompt, and a fix for its cloning notebook is still open on Canopy's checklist. Bland Speech makes an instant clone from about 10 seconds of audio and a professional clone from 30 minutes or more, with consent confirmed for every voice.
Is Orpheus TTS still being updated?
The repository's last commit is dated 5 December 2025, it has no tagged releases, and Canopy's site describes its new speech models as private. Bland Speech is a current product with a v3 model, and new accounts get 50,000 characters free.
Can Bland Speech replace an OpenAI text to speech integration?
Yes. POST /v2/audio/speech follows OpenAI's audio speech contract, so a team points its OpenAI client at Bland's base URL, uses a Bland API key, and sets the model to btts-3.
