The open 82M TTS model alternatives in 2026 for cloning and more voices
The open 82M TTS model alternatives compared: open models and hosted APIs with voice cloning, more voices, and more languages, with 2026 pricing. ElevenLabs leads on natural English, Cartesia on latency, Chatterbox is the open source clone, and PyAI adds free cloning with no GPU to run.

On this page
Why developers look for the open 82M TTS model alternatives
The open 82M TTS model is a favorite open model, lightweight, fast and free, running even on a CPU. Developers look elsewhere when they hit one of its ceilings, not because it is slow.
Preset voices, no clones
The open 82M TTS model uses preset voices only and cannot clone a voice, since it was trained on under 100 hours of audio. A branded or personal voice needs another model or a hosted API.
Limited voices and range
About 54 preset voices suit clean narration, but lack the expressive and emotional range of larger models and commercial services.
Uneven languages
Version 1.0 added Spanish, French, Hindi, Japanese, Mandarin and a few more, yet weak grapheme to phoneme handling makes names, numbers and non English text the usual failure points.
You host it
The weights are free, but you own the GPU or CPU, the setup and the operations, which is a hurdle for teams that want a plug in API.
The open 82M TTS model alternatives at a glance
The open 82M TTS model replacements sorted by hosted price, cloning support and the workload each one suits, with vendor rates rechecked on 30 October 2026.
| Tool | Built for | Features | Free option | Price (2026) | Cloning |
|---|---|---|---|---|---|
| PyAI | Developers who want a hosted, reliable voice API | Speak and Cast voices, Hear transcription, the Omni speech to speech model, an OpenAI compatible API, and WAV, PCM and G.711 formats | $50 credit, no credit cardSandbox keys never bill | $0.04/min Speak, $0.02/min Cast | Free |
| ElevenLabs | Narrators producing long form audio | A large voice catalog with cloning, dubbing, and studio tools for long form narration | Free plan | $6 Starter. Flash $0.04, Multilingual $0.08 per 1K chars | Yes |
| Cartesia | Builders of low latency voice agents | The Sonic model for fast streaming speech, with voice cloning for interactive agents | 20,000 credits | $5 Pro, $49 Startup | Yes |
| Fish Audio | Creators wanting low cost cloning | Community voice models with cloning on paid plans, aimed at individual creators | Free credits, personal use | Plus $11 a month billed yearly | Yes, on paid plans |
| Deepgram | Developers generating speech in bulk | Aura text to speech alongside Nova transcription, built for high volume developer apps | $200 credit | Aura-2 $0.030 per 1K chars | No |
| Chatterbox | Teams self hosting a cloning model | An open cloning model under the MIT license that you run on your own GPU | Free, MIT license | $0 plus your GPU | Yes, self hosted |
| Speechmatics | Teams wanting a low cost TTS API | A text to speech API alongside its broad language transcription | First 1M characters | $0.011 per 1K chars | No |
| the open 82M TTS model (for comparison) | Developers self hosting a light TTS model | A lightweight open text to speech model under the Apache license that runs on your own GPU | Free, Apache 2.0Runs on your own GPU | Free, self host, or under $1 per 1M chars hosted | No |
The open 82M TTS model cost vs the hosted alternatives
Self hosting the open 82M TTS model costs nothing in license fees, so the real comparison is against hosted engines that add cloning. Bars below price one million characters, roughly 1,111 minutes of audio at 900 characters a minute, on each vendor's published rate.
The open 82M TTS model alternatives, one million characters
A million characters on PyAI Speak lands near $44, cloning included
Bar length tracks the hosted bill for one million characters. The third column flags where each engine is strongest against the open 82M TTS model.
- Speechmatics TTSPer character APIFirst 1M free
$11 per million characters on Speechmatics, whose first million characters cost nothing
- Fish AudioCreator cloningBudget cloning
About $15 per million characters on Fish Audio S2 Pro, billed per million UTF-8 bytes
- Deepgram Aura-2Developer TTSHigh volume
$30 per million characters for Aura-2 pay as you go, dropping to $27 on Growth
- PyAI SpeakReliable API, free cloningFree cloning
About $44 per million characters for Speak at four cents a minute, with cloned voices billed at the same rate, while Cast narration costs half
- ElevenLabsPremium narrationMost natural
From $40 on Flash to $80 per million on Multilingual for Flash at the low end and Multilingual v2 or v3 at the high end
- the open 82M TTS modelOpen modelFree, self host
Free to self host. Under $1 per million characters on a hosted endpoint, with preset voices and no cloning
Vendor rates rechecked 30 October 2026. Third party the open 82M TTS model endpoints were quoted under $1 per million characters in April 2025.
The open 82M TTS model competitors reviewed
Each of the seven answers a specific the open 82M TTS model ceiling: PyAI for hosted cloning without a GPU, ElevenLabs for English realism, Cartesia for streaming speed, Fish Audio for budget clones, Deepgram for bulk app speech, Chatterbox for local cloning, and Speechmatics for the lowest per character bill.
PyAI
Built for a hosted the open 82M TTS model replacement that clones voices at no extra fee
PyAI bundles Speak for speech, Cast for long narration, Hear for transcription and Omni for phone agents under a single API key. Against the open 82M TTS model it swaps a self run model for a hosted endpoint with 200 voices, free voice cloning plus custom voices generated from a typed description. PyAI does not ship downloadable weights, so offline or air gapped the open 82M TTS model builds are better served by Chatterbox.
Speak costs $0.04 a minute while Cast, aimed at audiobooks and hour long videos, costs $0.02 a minute. Audio starts in about 53 ms warm and 98 ms cold in PyAI's own tests, and output comes back as WAV, PCM or G.711. Cloned voices stay on the account with no monthly fee attached.
PyAI is recent, so it has no G2 reviews yet.
PyAI features
- Speak
- Hosted speech with 200 stock voices behind an OpenAI style endpoint, so the open 82M TTS model code that already calls a speech route needs few changes.
- Cast
- Chapter length narration rendered in the background at $0.02 a minute, half the Speak rate.
- Voice cloning
- Enrollment from a short clip costs nothing. Generation with the clone bills at the normal Speak rate.
- Commercial use
- Commercial rights come with the base rate rather than a pricier tier.
PyAI pricing
Speak
$0.04/min
Voiceovers
Cast
$0.02/min
Long narration
Cloning
Free
Short clip
Usage bills by the minute, test keys stay free, and a fresh account opens with $50 of free credit, no credit card needed. Sample the stock voices on PyAI Speak.
PyAI pros and cons
Pros
- Commercial rights and cloning sit inside the base $0.04 a minute, so a replacement for the open 82M TTS model adds no seat or tier fee.
- First audio arrives in about 53 ms warm in PyAI's tests, quick enough to reuse the same voice later on Omni phone calls.
- Output comes back as WAV, PCM or G.711 at the requested sample rate, which matched in 100% of PyAI's format tests.
- Per minute billing with no subscription means a month of light the open 82M TTS model style batch jobs costs only the minutes rendered.
Cons
- No offline or downloadable model, so air gapped builds that run the open 82M TTS model locally today cannot move to PyAI.
- No public G2 or Capterra ratings yet, and every latency figure on this card comes from PyAI's own benchmark runs.
PyAI reviews
No PyAI listing exists on G2 or Capterra yet, so this card shows no ratings. The fair check is to clone one voice, render the same the open 82M TTS model script on both, and compare the files side by side.
Choose PyAI if you like the open 82M TTS model's output but need branded clones and would rather not keep a GPU box running.
ElevenLabs
Built for expressive English clones for long reads, hosted
ElevenLabs is the hosted engine most the open 82M TTS model comparisons rate above it for English realism, with instant and professional clones and a large shared voice library for long reads.
A free plan comes first, then Starter at $6 and Creator at $22 a month. On the API, Flash costs $0.04 and Multilingual $0.08 per 1,000 characters. G2 users give it 4.5 out of 5. The tradeoff against the open 82M TTS model is a metered hosted bill and voices that live inside the account rather than free output on your own GPU. ElevenLabs fits a developer using the open 82M TTS model who has outgrown preset voices and wants studio grade English narration or branded clones without running inference.
ElevenLabs features
- Instant and pro clones
- ElevenLabs offers instant clones from a short sample and professional clones trained on more audio.
- Shared voice library
- A large shared ElevenLabs library supplies ready voices for long reads such as audiobook chapters.
- Flash and Turbo
- ElevenLabs Flash and Turbo cost $0.04 per 1,000 characters on the API, suited to drafts and quick renders.
- Multilingual v2 and v3
- Multilingual v2 and v3 run $0.08 per 1,000 characters for final, non English or expressive output.
- Plan ladder
- ElevenLabs starts free, then Starter costs $6 and Creator $22 a month.
Against the open 82M TTS model's pain points: ElevenLabs
ElevenLabs is the hosted step up from the open 82M TTS model for clones and English realism, at a per character cost the open 82M TTS model never had.
- Voice cloning: ElevenLabs offers both instant and professional clones, the capability the open 82M TTS model lacks because of its small training set
- Voice range: Most the open 82M TTS model comparisons rate ElevenLabs English delivery as more expressive across long chapters than the open 82M TTS model's presets
- Languages: ElevenLabs Multilingual v2 and v3 cover non English scripts, though G2 reviewers still report pronunciation slips on names
- Hosting: ElevenLabs runs fully hosted, so the GPU, setup and uptime work of self hosted the open 82M TTS model disappears
- Phone calls: ElevenLabs telephony and agent pricing were not part of the sources checked here, so confirm call support before relying on it
- Volume cost: One million characters costs about $40 on ElevenLabs Flash or $80 on Multilingual, against hardware only for the open 82M TTS model
ElevenLabs pros and cons
Pros
- Most the open 82M TTS model comparisons rate its English delivery as more expressive, especially across long audiobook chapters.
- Flash at $0.04 per 1,000 characters keeps the cheaper tier usable for drafts, then Multilingual for final renders.
- G2 rates it 4.5 across 1,260 reviews, a large review base of any engine on this list.
Cons
- Multilingual at $0.08 per 1,000 characters makes a million characters about $80, versus near zero on self hosted the open 82M TTS model.
- G2 reviewers flag monthly credits that run out fast, plus pronunciation slips on names, much like the open 82M TTS model's own weak spot.
- Creator at $22 a month is the usual floor for serious clone work, a fixed cost the open 82M TTS model never had.
ElevenLabs reviews
"The absolute best part is just how genuinely human the voices sound: the natural pauses, the breathing, and the way emphasis shifts."
"The main pain point is how it handles dense legal terminology, citations and Latin maxims. The model frequently stumbles over citations."
"The best part about ElevenLabs is the free credit thing they have."
Choose ElevenLabs if you run a narration project using the open 82M TTS model that needs expressive English clones and budget is secondary.
Cartesia
Built for streaming speech that feeds live voice agents
Cartesia builds its Sonic models for streaming, with first audio claimed at 82 to 100 ms network included, and sells an agent platform beside the speech API.
The free tier holds 20K credits. Pro at $5 adds instant cloning, Startup at $49 brings 1.25M credits, and Scale is $299. Agents bill $0.03 a minute, plus $0.014 on a Cartesia number, with the LLM billed separately. Cartesia rewards a build that uses the open 82M TTS model moving into live conversation, where first audio speed decides whether a caller hears a natural pause or dead air. The cost is a smaller shared voice library than ElevenLabs and a credit model that steps up quickly from the free 20K tier, so steady production work lands on Startup or Scale rather than the free plan.
Cartesia features
- Sonic streaming
- Cartesia builds Sonic models for streaming, claiming first audio in 82 to 100 ms with network time included.
- Pro plan clones
- Instant cloning opens on the Cartesia Pro plan at $5 a month with 100K credits.
- Startup plan
- Cartesia Startup costs $49 a month for 1.25M credits and 2 professional clone slots.
- Scale plan
- Cartesia Scale is $299 a month with 8M credits for heavier production use.
- Agent platform
- Cartesia voice agents bill $0.03 a minute, with the LLM charged separately.
- Phone numbers
- A Cartesia number adds $0.014 a minute to agent calls.
Against the open 82M TTS model's pain points: Cartesia
Cartesia moves a developer using the open 82M TTS model onto hosted streaming speech with cheap cloning and a route into phone agents.
- Voice cloning: Cartesia opens instant cloning at the $5 Pro tier, while trained professional voices wait for the $49 Startup plan
- Voice range: Cartesia's expressive range versus the open 82M TTS model was not measured in the sources checked, so audition Sonic voices on your own script
- Languages: Cartesia language coverage was not listed in the pricing checked for this page, so confirm support for non English scripts
- Hosting: Cartesia hosts Sonic as an API, so a team using the open 82M TTS model stops maintaining its own GPU or CPU server
- Phone calls: Cartesia's agent platform handles live calls for $0.03 a minute, plus $0.014 on a Cartesia number, where the open 82M TTS model only narrates in batch
- Volume cost: Cartesia Startup buys 1.25M credits for $49 a month, a fixed monthly bill in place of the open 82M TTS model's hardware cost
Cartesia pros and cons
Pros
- Instant cloning starts on the $5 Pro plan with 100K credits, the cheapest hosted clone tier reviewed here.
- Agents at $0.03 a minute plus $0.014 on a Cartesia number make it a direct path from narration into phone calls.
Cons
- Speed claims of 82 to 100 ms are vendor numbers, so time your own the open 82M TTS model script before trusting them.
- Professional clones need Startup at $49 for 2 slots or Scale at $299 for 4.
- Only 43 G2 reviews exist, so the 4.3 rating rests on a thin sample.
Cartesia reviews
"The biggest advantage is the low latency, which makes conversations feel more natural and responsive, and the voice quality is quite natural, especially for conversational use cases."
"Some areas could provide more detailed information around individual calls. I would also like more flexibility and deeper reporting as the number of agents and calls grows."
"The best part is Cartesia provides a free trial too, so I try such free alternatives before purchasing any plan."
Choose Cartesia if you are pushing the open 82M TTS model into live conversation and first audio speed decides the build.
Fish Audio
Built for cheap hosted clones and a community voice library
Fish Audio pairs a cheap hosted clone with a public community voice library, so a developer using the open 82M TTS model can borrow or upload a voice instead of settling for presets.
The free plan gives 8,000 monthly credits for personal work. Plus is $11 a month billed yearly or $15 monthly. The S2 Pro API charges $15 per million UTF-8 bytes, and non English text uses more bytes per character. For a side project using the open 82M TTS model the appeal is a cloned voice for a few dollars a month, pulled from or added to the community library instead of built from scratch. The tradeoff is that quality and licensing of community voices vary, and the S2 Pro API bills per UTF-8 byte rather than per character, so non Latin scripts cost more per spoken minute.
Fish Audio features
- Community library
- Fish Audio hosts a public community voice library, so a voice can be borrowed instead of built.
- Free cloning
- Fish Audio lets free accounts clone voices, with 3 public slots to keep them in.
- Free credits
- Fish Audio gives 8,000 credits a month free, up to 7 minutes of generation for personal use.
- Plus and Pro
- Fish Audio Plus costs $15 month to month or $11 on annual billing, with Pro at $100 monthly or $75 annually.
- S2 Pro API
- The Fish Audio s2-pro and s2.1-pro API models charge $15 per million UTF-8 bytes.
Against the open 82M TTS model's pain points: Fish Audio
Fish Audio gives a developer using the open 82M TTS model cloning and a shared voice library at a budget hosted price.
- Voice cloning: Fish Audio lets even free users clone, with 3 public voice slots, so a custom voice costs nothing to test
- Voice range: Fish Audio's community library offers far more ready voices than the open 82M TTS model's 54 presets, for variety without training
- Languages: Fish Audio bills by UTF-8 byte, so non English scripts use more bytes per character and cost more than English
- Hosting: Fish Audio runs the model in its cloud, so the GPU and operations burden of the open 82M TTS model moves to the vendor
- Phone calls: Fish Audio telephony or agent support was not covered in the sources checked, so treat Fish as a narration engine like the open 82M TTS model
- Volume cost: Fish Audio's $15 per million bytes is roughly $15 per million English characters, plus Plus at $15 monthly for client work
Fish Audio pros and cons
Pros
- The S2 Pro API at $15 per million bytes is the cheapest hosted cloning option against the open 82M TTS model on this page.
- Even the free plan includes enhanced cloning with 3 public voice slots, so testing a clone costs nothing.
- A public community library offers ready voices that the open 82M TTS model's 54 presets cannot match for variety.
Cons
- The FAQ limits free plan audio to personal, non commercial projects, so client work needs Plus at $15 billed monthly.
- Byte based API billing means non English scripts cost more per character than the English rate suggests.
Fish Audio reviews
"I like the smoothness and speed of generation, the quality of the audio, and the very cheap price, especially with the S2 model."
Choose Fish Audio if you want a cloned voice for a side project using the open 82M TTS model at a few dollars a month.
Deepgram
Built for high volume speech alongside transcription
Deepgram sells Aura-2 speech next to Nova-3 transcription and a Voice Agent API, which suits the open 82M TTS model users whose app also needs to listen.
Aura-2 runs $0.030 per 1,000 characters on pay as you go, and the agent API lists $0.075 a minute. Sign up brings $200 that never expires, with no credit card needed. Deepgram fits a developer using the open 82M TTS model whose app also has to listen, since one account covers Aura-2 speech, Nova-3 transcription and the Voice Agent API on a single bill. The catch for a pure text to speech need is a narrower voice range than a dedicated engine such as ElevenLabs, and no voice cloning, so branded voices are off the table.
Deepgram features
- Aura-2 voices
- Deepgram Aura-2 charges three cents per 1,000 characters, falling to $0.027 for Growth customers.
- Nova-3 transcription
- Nova-3 speech to text sits on the same Deepgram account for apps that also listen.
- Voice Agent API
- The Deepgram Voice Agent API lists $0.075 a minute on pay as you go and $0.068 on Growth.
- Lasting credit
- $200 of Deepgram credit never expires and needs no credit card.
Against the open 82M TTS model's pain points: Deepgram
Deepgram solves the open 82M TTS model's hosting and call gaps for developers, though it lists no cloning.
- Voice cloning: Deepgram lists no voice cloning on its pricing page, so a custom voice still needs another engine, as with the open 82M TTS model
- Voice range: Aura-2's expressive range compared with the open 82M TTS model was not measured in the sources checked, so test emotional lines before switching
- Languages: Aura-2 language coverage was not listed in the pricing checked for this page, so verify each target language with Deepgram
- Hosting: Deepgram hosts Aura-2 as an API, though no studio editor exists, so non technical users rely on developers
- Phone calls: Deepgram's Voice Agent API at $0.075 a minute adds the turn taking the open 82M TTS model lacks for live phone agents
- Volume cost: Aura-2 at $0.030 per 1,000 characters makes a million characters $30, and the $200 credit covers about 6.7 million
Deepgram pros and cons
Pros
- $200 of sign up credit never expires and covers roughly 6.7 million Aura-2 characters at $0.030 per 1,000.
- Growth pricing drops Aura-2 to $0.027 per 1,000 characters for teams with steady app traffic.
Cons
- The pricing page lists no voice cloning, so it fixes the open 82M TTS model's hosting problem but not its preset voice limit.
- Built as a developer API with no studio editor, which slows non technical teams used to the open 82M TTS model web demos.
Deepgram reviews
"Deepgram is really a developer-friendly solution and is extremely fast if you compare it against other vendors like OpenAI Whisper and Speechmatics."
"Although the accuracy is very good, it can sometimes be a bit difficult to adjust the models for very specific cases or when there is technical jargon."
"It supports word recognition and speaker diarization, and it also has a very good price offering comparatively."
Choose Deepgram if you are replacing the open 82M TTS model in an app that also transcribes users and want one bill for both directions.
Chatterbox
Built for free, open source voice cloning
Chatterbox from Resemble AI is the local model the open 82M TTS model users try first for cloning. The MIT licensed family includes a 350M Turbo with tags like laugh, a 110M Nano for CPUs, and a 500M multilingual model.
Weights cost nothing, as with the open 82M TTS model, so the bill is hardware and setup. Every clip carries a Perth neural watermark, and the flow-matching voice-cloning model and the second-generation multilingual voice-cloning model are the usual open cloning rivals. Chatterbox answers the same brief as the open 82M TTS model, a local model with free weights, while adding the cloning the open 82M TTS model lacks, which is why self hosting teams try it first. The real cost is operational: a GPU to serve the 350M or 500M models, setup time, and a Perth watermark on every clip, so it rewards teams that already run their own inference.
Chatterbox features
- MIT license
- Chatterbox from Resemble AI ships under the MIT license with free weights.
- Turbo 350M
- Chatterbox Turbo is a 350M English model that accepts paralinguistic tags such as laugh.
- Nano 110M
- Chatterbox Nano at 110M runs on CPU at about 3x realtime on 8 cores.
- Multilingual V3
- A 500M Chatterbox Multilingual V3 model handles languages beyond English.
- Clip cloning
- Chatterbox clones a voice from a reference clip of about 10 seconds.
- Perth watermark
- Every Chatterbox output carries the Perth neural watermark.
Against the open 82M TTS model's pain points: Chatterbox
Chatterbox keeps the open 82M TTS model's free, local model while adding the cloning the open 82M TTS model cannot do.
- Voice cloning: Chatterbox clones from a 10 second reference clip on your own hardware, the main ability the open 82M TTS model is missing
- Voice range: Chatterbox Turbo accepts tags such as laugh, adding expression that the open 82M TTS model's preset voices cannot produce
- Languages: Chatterbox Turbo and Nano are English only, so non English work needs the 500M Multilingual V3 model and more GPU memory
- Hosting: Chatterbox comes with no hosted endpoint or vendor support, so uptime and scaling stay with your team, as with the open 82M TTS model
- Phone calls: Chatterbox is a speech model without telephony, so live calls still need a separate agent and phone stack
- Volume cost: Chatterbox weights are free, so the cost at volume is your hardware, with the flow-matching voice-cloning model and the second-generation multilingual voice-cloning model as other open rivals
Chatterbox pros and cons
Pros
- MIT license and free weights keep the zero license cost that drew teams to the open 82M TTS model in the first place.
- Clones from a roughly 10 second reference clip, the main ability the open 82M TTS model is missing.
- Chatterbox Nano at 110M reaches about 3x realtime on 8 CPU cores, close to the open 82M TTS model's CPU friendly footprint.
Cons
- Turbo and the original model are English only. Other languages need the larger 500M multilingual build and more GPU memory.
- No hosted endpoint or vendor support comes with the open model, so uptime and scaling stay your job, as with the open 82M TTS model.
- The Perth watermark is embedded in every clip, which some commercial users will need to clear with their own policy.
- Chatterbox generates speech only, with no telephony, so live calls still need a separate agent and phone stack.
Choose Chatterbox if you need cloning to stay on local hardware and a model with zero licensing fees, as with the open 82M TTS model, matters.
Speechmatics
Built for low cost hosted speech billed per character
Speechmatics made its name in transcription and added a speech API that draws from the same credit balance, which keeps billing simple.
Generation costs $0.011 per 1,000 characters after a free first million, about 20 hours of audio. The pricing page lists no cloning. Speechmatics fits a workload using the open 82M TTS model moving to hosted speech purely to drop the GPU, where one credit balance for transcription and generation keeps billing simple. The limits are real: the pricing page lists no cloning and the preset voice range is narrow, so branded or custom voices need ElevenLabs or Fish Audio, and a team that needs expressive delivery will outgrow it.
Speechmatics features
- Per character rate
- Speechmatics charges $0.011 per 1,000 characters for speech generation, $11 per million.
- Free million
- The first million Speechmatics characters are free, about 20 hours of audio.
- Shared balance
- Speechmatics speech to text and speech generation draw from one credit balance.
- Review base
- Speechmatics carries 4.7 on G2 across 75 reviews, earned mainly by its transcription work.
Against the open 82M TTS model's pain points: Speechmatics
Speechmatics is the lowest hosted bill against the open 82M TTS model on this page, without cloning.
- Voice cloning: Speechmatics lists no voice cloning on its pricing page, so a branded voice needs another engine
- Voice range: Speechmatics added speech generation recently and offers a smaller catalog than ElevenLabs, so preview the voices first
- Languages: Speechmatics language support for speech generation was not listed in the pricing checked, so confirm before moving non English work
- Hosting: Speechmatics hosts generation on the same account as transcription, removing the open 82M TTS model's server setup
- Phone calls: Speechmatics telephony or agent tooling was not covered in the sources checked, so live calls need another vendor
- Volume cost: At $0.011 per 1,000 characters, 10 million Speechmatics characters cost $110, with the first million free
Speechmatics pros and cons
Pros
- $0.011 per 1,000 characters means a million characters costs $11, the lowest hosted bill reviewed here.
- One credit balance covers speech to text and speech generation, handy for apps that both listen and speak.
- The first million characters are free, roughly 20 hours of audio, enough to test a full the open 82M TTS model backlog.
- G2 rates Speechmatics 4.7 from 75 reviews, mostly for transcription accuracy.
Cons
- No voice cloning appears on the pricing page, so branded voices still need another engine.
- Reviewers know it for transcription. Speechmatics speech generation has a short track record and fewer voices than ElevenLabs.
- Speechmatics does not list language support for speech generation on its pricing page, so confirm coverage before moving non English work.
Speechmatics reviews
"We appreciate the relative ease of integrating Speechmatics into our broadcast monitoring workflow. Customer support is reliable and provides quick responses."
"It appears to fill in a whole lot of nonsense when it hears noise or background chatter, creating disconcerting false positive transcriptions."
"Functionalities and pricing tailored to your exact application and business domain, a good software stack and API, reliable transcoding."
Choose Speechmatics if you are moving a workload using the open 82M TTS model to hosted speech purely on cost and preset voices are acceptable.
The open 82M TTS model alternatives by job and use case
Match the open 82M TTS model workload you are replacing to the engine built for it.
Which engine suits each workload?
| Job | Pick | Why |
|---|---|---|
| Natural voiceovers | PyAI Speak | Free cloning, $0.04 a minute |
| Audiobooks and long narration | PyAI Cast, ElevenLabs | Cast at $0.02/min, or ElevenLabs for expressive clones |
| Voice agents | Cartesia, PyAI Omni | Low latency streaming |
| App speech at volume | Deepgram | Aura-2 at $30 per million, $200 trial credit |
| Local cloning | Chatterbox, the flow-matching voice-cloning model | MIT or open weights, clone from a 10 second clip |
| Cheapest per character | Speechmatics | First 1M characters free, then $11 per million |
Text to speech by price
| Tool | Price |
|---|---|
| Speechmatics TTS | $0.011 per 1K characters |
| Fish Audio | $15 per million bytes on S2 Pro |
| PyAI Cast | $0.02/min narration |
| Deepgram Aura-2 | $0.030 per 1K characters |
| ElevenLabs Flash | $0.04 per 1K characters |
Open source voice cloning
| Model | License | Note |
|---|---|---|
| Chatterbox | MIT | Turbo, Nano and multilingual builds, watermarked output |
| the second-generation multilingual voice-cloning model | non-commercial model license (non-commercial) | Clones from about 6 seconds of reference audio. Commercial use needs a separate license, so check before shipping |
| the flow-matching voice-cloning model | Pretrained weights CC-BY-NC (non-commercial) | Flow matching, zero shot cloning at about 336M. The CC-BY-NC weights block commercial use, so verify the license first |
| Higgs Audio V2 | Open weights | Expressive multilingual cloning, needs about 24GB VRAM |
Free and open source alternatives to the open 82M TTS model
To keep the open 82M TTS model's zero license cost and add cloning, Chatterbox under MIT runs locally on your own GPU, while the flow-matching voice-cloning model and the second-generation multilingual voice-cloning model also run locally but ship pretrained weights under non-commercial licenses, so confirm commercial rights before you rely on them, and Dia from Nari Labs covers two speaker dialogue without cloning. For a hosted test before committing, trial balances differ widely: PyAI gives $50 with no credit card, Cartesia 20K free credits, Deepgram $200 that never expires, and Speechmatics one million free characters.
When the open 82M TTS model is still the right pick
The open 82M TTS model still wins for batch English narration on a laptop CPU, in offline builds, and anywhere an 82M model with no per call fee beats a hosted bill. The open 82M TTS model's sweet spot is 100 to 200 token passages. Move only when clones, broader voice choice or zero GPU upkeep become requirements.
How to move beyond the open 82M TTS model in 4 steps
Most the open 82M TTS model scripts move in an afternoon, since the change is one endpoint and a voice ID.
Step 1 of 4
Pick an API
Shortlist a hosted engine whose request shape mirrors your the open 82M TTS model wrapper, so only the URL and key change.
Step 2 of 4
Clone a voice
Record about 10 seconds of quiet, single speaker audio and enroll it, since the open 82M TTS model has no clone to carry over.
Step 3 of 4
Compare on your script
Render the same passage on the open 82M TTS model and the new engine, then compare names, numbers and first audio time.
Step 4 of 4
Switch one workload
Cut over one product feature first and keep the open 82M TTS model as a local fallback until the hosted output holds up.
Which alternative to the open 82M TTS model to pick
Pick the open 82M TTS model replacement by the ceiling you hit. Local cloning points to Chatterbox or the flow-matching voice-cloning model. English realism to ElevenLabs. Streaming speed to Cartesia. PyAI Speak suits teams that want clones without GPU upkeep at $0.04 a minute, and a test account opens with $50 and no credit card.
More text to speech comparisons
The open 82M TTS model alternatives FAQs
Common questions from developers comparing the open 82M TTS model.
Does the open 82M TTS model support voice cloning?
No. the open 82M TTS model is a small open model with preset voices only and cannot clone a voice, because it was trained on under 100 hours of audio. Local cloning comes from Chatterbox or the flow-matching voice-cloning model, and hosted clones from PyAI, ElevenLabs or Fish Audio.
What is a good alternative to the open 82M TTS model?
Chatterbox for local use, PyAI or ElevenLabs hosted. Chatterbox keeps an MIT license and adds cloning on your own GPU. PyAI Speak clones at no extra fee for $0.04 a minute with no hardware to run, while ElevenLabs costs more but rates highest for English realism.
What does running the open 82M TTS model cost?
the open 82M TTS model is free and open source. the weights for the open 82M TTS model are Apache 2.0 with no license fee, so the cost is your own GPU or CPU and the setup time. Hosted endpoints for the open 82M TTS model run under $1 per million characters, near $0.65 to $0.80, yet keep the same preset voices. Hosted engines with cloning start near $15 per million on Fish Audio.
How many voices and languages does the open 82M TTS model have?
About 54 preset voices. the open 82M TTS model v1.0 covers American and British English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese and Mandarin, though several of those get only one or two voices and none can be cloned. The 82 million parameter model is Apache 2.0 and runs on a CPU, so the ceiling is voice range, not speed.
What is a alternative to the open 82M TTS model with voice cloning?
Chatterbox, Fish Audio, PyAI and ElevenLabs. Chatterbox clones locally from a roughly 10 second clip under MIT. Fish Audio Plus costs $11 a month billed yearly, PyAI enrolls clones free and bills $0.04 a minute, and ElevenLabs suits expressive English clones at a higher rate.
Is there a better open source TTS than the open 82M TTS model?
Better at cloning, not at size. the open 82M TTS model stays the lightest at 82M parameters. Chatterbox Nano at 110M also runs on a CPU and adds cloning, the flow-matching voice-cloning model clones zero shot at about 336M, and Dia at 1.6B handles dialogue, so each step up costs more hardware.
Test PyAI on your own audio and calls.
Get a sandbox key and hear the output in under a minute. Sandbox keys are free, with $50 in starter credit and no credit card required. Speak is $0.04/min, Cast $0.02/min, Hear $0.001/min and Omni $0.05/min + $0.01 telephony.