Piper TTS is a free, open-source neural text-to-speech engine that converts text into speech entirely on your own device, with no cloud API call required. Originally built inside the Rhasspy voice-assistant project, it now ships under the Open Home Foundation as a fast, lightweight engine tuned for edge hardware like a Raspberry Pi.
You will run into Piper anywhere someone needs speech synthesis without a network round trip: home automation assistants, screen readers, offline kiosks, and cost-sensitive IVR prototypes. This entry covers how it works, what its licensing actually says, the metrics that describe a voice, how it stacks up against hosted APIs, and when it is and is not the right call.
What Is Piper TTS?
Piper TTS is a local, neural text-to-speech engine: you install it on your own hardware, feed it text, and it returns audio without sending anything to a third-party server. That single property, running on-device, is what separates it from every hosted TTS API on the market. In practice that is one command: pip install piper-tts pulls the engine, and you download voice files separately.
The project started inside Rhasspy, an open-source offline voice-assistant toolkit. In October 2025, the original repository, rhasspy/piper, was archived and made read-only under its MIT license. Active development now lives at OHF-Voice/piper1-gpl, maintained by the Open Home Foundation, the nonprofit that also builds Home Assistant's voice stack. The repository shipped a release as recently as August 2026 and carries over 5,000 GitHub stars. The star counts mislead here: the archived repository still shows over 11,000 stars against the active one's 5,284, purely because it had a two-year head start. Releases, not stars, tell you where the work is happening. This is not an abandoned fork; it is where the project actually lives now.
Piper is not a research paper or a novel architecture. It is an engineering package around an existing model family, built for one job: get intelligible, natural-sounding speech out of modest hardware, fast, for free.
How Piper TTS Works
Piper builds each voice on VITS, a neural architecture that generates speech directly from text without a separate vocoder step. The trained model exports to ONNX, a portable format that runs on CPU without a GPU or specialized inference server.
Before the neural model ever runs, Piper converts your input text into phonemes, the individual speech sounds a language uses. It does this with espeak-ng, a rule-based phonemizer, rather than learning that step end-to-end. Splitting the pipeline this way keeps the neural model smaller and faster, at the cost of inheriting espeak-ng's occasional mispronunciations on unusual words.
Each voice ships as two files: an .onnx model and a .onnx.json configuration file. You download a voice, point Piper at it, and it synthesizes audio locally, whether that's on a server, a desktop, or a Raspberry Pi.
Licensing: What Changed and What It Means
Piper's licensing has two separate layers, and both changed with the move to its new home.
The engine's own code shifted license. The archived rhasspy/piper repository was MIT-licensed. The current repository, OHF-Voice/piper1-gpl, ships under GPL-3.0, confirmed directly in its repository metadata. The engine also embeds espeak-ng for phonemization, and espeak-ng is itself GPL-licensed software.
Voice models are licensed separately from the engine, and that license varies by voice. Piper's own documentation states the project is intended for personal use and text-to-speech research only, adds no extra restrictions of its own on voice models, and warns that some individual voices carry more restrictive licenses. Every voice ships a MODEL_CARD file, and the project tells you to check it before you rely on that voice commercially.
Practically, a licensing review for Piper has two steps. First, work out what GPL-3.0 costs you: it is a copyleft license, so distributing an application that links the engine can oblige you to release your own source under compatible terms, which is the substantive change from the old MIT repository and the reason the move matters commercially. Running it as an internal service you never distribute is a different question from shipping it inside a product. Pipecat's Piper integration docs make the same point and recommend running Piper as a separate HTTP server rather than linking it in, which keeps the engine at arm's length from your own source. Second, check the specific MODEL_CARD for each voice you actually ship.
The Metrics That Describe a Piper TTS Voice
Four numbers describe what you are actually getting from a given Piper TTS voice.
Quality tier and sample rate. Piper publishes voices in four tiers, x_low, low, medium, and high, which trade model size and inference speed for a higher sample rate and cleaner audio: x_low and low voices render at 16 kHz, medium and high at 22.05 kHz. The spread is lopsided: of the 175 voices in Piper's official voice index, 121 are medium, 26 low, and 14 each x_low and high.
Model size. Piper's voice index publishes the byte size of every model file. The x_low tier runs about 21 to 28 MB, low and medium cluster around 63 MB, and high-quality voices run from 63 MB to roughly 137 MB. Even the largest is small enough to bundle with an application rather than stream from a server.
Real-time factor (RTF). RTF measures synthesis speed against playback time: an RTF of 0.5 means the engine generates ten seconds of audio in five seconds. A December 2025 arXiv study on TTS phonemization used a Persian Piper voice as its baseline and measured an RTF of 0.153 on a 12th Gen Intel Core i7-1255U running CPU-only inference, roughly six and a half times faster than real time. The same paper puts Piper's general RTF at about 0.2. Treat either as one setup, not a universal number: yours depends on the voice's quality tier and the CPU you run it on.
Voice and language coverage. Piper's voice index ships 175 voices across 56 language and locale combinations, from Arabic and Welsh to Japanese, Korean, and simplified Chinese. Check the index rather than the docs page: the hand-maintained language list in Piper's documentation still shows 44 locales and misses a dozen that have shipping voices. Coverage is uneven, with more than twenty named English voices and exactly one for some locales.
None of these are marketing numbers. They are the specs you actually need to size a deployment: which quality tier fits your latency budget, how much disk or memory a voice needs, and whether your target language has a voice at all.
How Piper TTS Compares to Other TTS Engines
Piper's real competition isn't other open-source engines: it's the decision between self-hosting and paying for a hosted API. The table below places it against the engines people most often compare it to.
| Engine | Deployment | License / pricing | Best fit |
|---|---|---|---|
| Piper TTS | Self-hosted, on-device | Free, GPL-3.0 engine; per-voice licenses vary | Offline apps, privacy-sensitive workloads, high call volumes you don't want billed per character |
| Tortoise-TTS | Self-hosted | Free, Apache 2.0 | Higher-fidelity open-source voice cloning when GPU time and latency aren't constraints, though the project has had no commits since November 2024 |
| Inworld TTS | Hosted API | Paid, character-based | Real-time conversational agents where time-to-first-audio is the binding constraint |
| Amazon Polly | Hosted API (AWS) | Paid, $4 per 1M characters standard, $16 neural, $30 generative, $100 long-form | Teams already on AWS wanting a managed, no-ops endpoint |
| OpenAI TTS | Hosted API | Paid, character- or token-based by model | Broad multilingual input coverage inside an existing OpenAI stack |
| Deepgram TTS (Aura-2) | Hosted API | Paid, character-based | Regulated-industry scripts, like medical or legal terminology |
| ElevenLabs | Hosted API | Paid, credit-based | Highest perceived naturalness and expressive, cloned voices |
The honest gap: Piper is fast and free, but its voices sit below the top hosted APIs on expressiveness and emotional range. The same December 2025 arXiv study that measured Piper's RTF also ran a listening test: 16 native Persian speakers rated seven utterances on a 1 to 5 naturalness scale, and baseline Piper scored 2.41 against 4.21 for a recording of a real human voice. That is one small study on one language, not a verdict on every Piper voice, but it is the only published mean opinion score for Piper, and the direction matches what you hear. For deeper feature and pricing detail on the hosted engines in this table, Cekura's own roundup of TTS APIs for voice agents and its comparison of TTS voice models go further than a single table can.
When to Choose Piper TTS, and When Not To
Choose Piper TTS when cost per character would otherwise scale badly, when audio has to be generated without a network path, or when policy requires that text and voice data never leave your own infrastructure. It also fits when you already run self-hosted infrastructure and can absorb one more service to operate.
Skip Piper TTS when a customer-facing voice's naturalness is the priority, since hosted APIs like ElevenLabs still lead there. Skip it too if you don't want to manage GPU or CPU capacity, model files, and updates yourself. Guaranteed vendor support is another reason to skip it: Piper is community-maintained, and its own README currently asks for more maintainers to step forward. It also doesn't ship a voice-cloning workflow; each voice is a separately trained model file, not something you generate from a sample on demand.
Most teams end up running Piper for one workload and a hosted API for another, rather than choosing one engine for everything.
Where Cekura Fits: Validating TTS Output Quality
Picking a TTS engine, Piper or otherwise, only answers half the question. The other half is confirming the output actually sounds right once it's wired into your voice agent, and that it keeps sounding right after every prompt change or voice swap.
Cekura tests, monitors, and helps self-improve voice and chat agents in production, and TTS output is one layer of that. Cekura's snapshot testing for voice responses captures both the SSML markup and the resulting audio for a given voice configuration, so a regression in pauses, emphasis, pitch, or timing shows up as a diff before it reaches a caller, not after.
Cekura's own benchmarks score full voice stacks on a Voice Tone + Clarity metric, but they evaluate hosted platforms like ElevenLabs, Retell, and LiveKit end to end rather than testing individual TTS engines like Piper in isolation, so there's no Piper-specific benchmark figure to cite here. For scale, Cekura's benchmarks put ElevenLabs at 4.47 out of 5 on Voice Tone + Clarity across full-stack scored calls, which is the bar a hosted API clears when naturalness is the whole reason you are paying for it. That score and Piper's 2.41 are not comparable numbers: one comes from Cekura's evaluators scoring a complete hosted stack, the other from human raters judging a single Persian voice. What does transfer is the underlying discipline: whichever engine you pick, you still need a repeatable way to catch voice quality regressions before a caller hears one. If you want to see that run against your own stack, Cekura's team will set up a walkthrough using the agent and the TTS engine you have already picked.
Frequently Asked Questions
Is Piper TTS free to use commercially?
The engine itself is free and open-source under GPL-3.0. Whether a specific voice is free for commercial use depends on that voice, not the engine: Piper's documentation says voices are meant for personal use and research, some carry more restrictive terms, and it tells you to check each voice's MODEL_CARD file before shipping it commercially.
Does Piper TTS need an internet connection to work?
No. Piper synthesizes audio entirely on local hardware, from a Raspberry Pi to a full server, using an ONNX model file per voice. That's the core tradeoff against hosted APIs: no per-character billing or network latency, but you own the compute, updates, and uptime yourself.
How many languages does Piper TTS support?
Piper's voice index lists 175 voices across 56 language and locale combinations, each a separately trained model file. One caveat: the language list in Piper's documentation is hand-maintained and still shows 44 locales, so check the voice index for anything recent. Coverage per language varies, from more than twenty named English voices down to one.
Is Piper TTS as good as ElevenLabs or other hosted APIs?
Not on perceived naturalness. Piper's voices are clear and fast but sit below top hosted APIs like ElevenLabs on expressiveness and emotional range, which is the tradeoff for running free and offline. If naturalness is the priority for a customer-facing voice, a hosted API is usually the better fit; for cost, privacy, or offline constraints, that gap matters less.
Who maintains Piper TTS now?
The Open Home Foundation, the nonprofit behind Home Assistant's voice stack, maintains Piper's current repository, OHF-Voice/piper1-gpl. The original rhasspy/piper repository is archived and read-only. The project's own documentation says it is actively looking for more maintainers, worth knowing if you're evaluating its long-term support model.
