Picovoice alternatives in 2027 for on-device and cloud voice
Picovoice alternatives compared: on-device open models and cloud speech APIs for wake word, speech to text and voice agents, with 2026 pricing. Vosk and Whisper are the open on-device options, Deepgram and AssemblyAI lead on cloud transcription, and PyAI is the cloud voice API.

On this page
Why teams look for Picovoice alternatives
Picovoice packs wake word (Porcupine), streaming and batch speech to text (Cheetah, Leopard), text to speech (Orca) and intent (Rhino) into SDKs that run offline on the device. The 2026 searches for alternatives come from its free tier ending and its turn toward enterprise buyers, not from weak engines.
Free tier ended
Picovoice Free Tier AccessKeys stopped working after 30 June 2026, and Home Assistant users and hobby builders had no non-commercial plan to move to.
Sales led pricing
Paid plans now start with a sales call rather than a published rate, and the single G2 review of Picovoice, a 5.0, still names pricing as its one complaint.
Enterprise focus
Picovoice says it is concentrating on enterprise deployments, which leaves indie developers and small device makers without a self serve paid tier.
Cloud may be fine
Many teams chose Picovoice for cheap voice, not for an offline rule. When audio is allowed to leave the device, a per minute cloud API removes the SDK license entirely.
Picovoice alternatives at a glance
Rows compare what each option is built for, its free starting point, a batch price and where the audio is processed, using vendor pricing pages and Google's published docs, checked on 30 October 2026.
| Tool | Built for | Features | Free option | Price (2026) | Runs |
|---|---|---|---|---|---|
| PyAI | Teams wanting a cloud API | Hear transcription and Omni agents on one OpenAI-compatible API, with WAV, PCM and G.711 | $50 credit, no credit cardSandbox keys never bill | Hear $0.001/min, Omni $0.05/min + $0.01 telephony | Cloud |
| Vosk | Offline app developers | An open source engine under Apache that transcribes on the device | Free, Apache 2.0 | $0, runs on your device | On-device |
| Whisper | Open transcription users | Open transcription weights under MIT you run on your own hardware | Free, MIT | $0, your own GPU | Self hosted |
| AssemblyAI | Cloud accuracy seekers | Accurate cloud transcription with audio intelligence add ons | $50 credit | Universal-3.5 Pro $0.21 an hour | Cloud |
| Deepgram | Developers who want transcription and voices from one API | Cloud transcription and voices across the Nova and Aura models | $200 credit | Nova-3 batch $0.0043/min | Cloud |
| Google Speech to Text | GCP based teams | Cloud transcription API managed through a Google Cloud account | 60 free minutes/mo | STT from $0.016/min | Cloud |
| Speechmatics | Accent heavy workloads | Cloud transcription tuned for a wide range of accents | $100 credit | Melia 1 batch $0.004/min | Cloud |
| Picovoice (for comparison) | On device voice teams | On device wake word, transcription and speech, quoted by sales | Free trial, no credit cardPaid plans by sales | Quoted, per device or enterprise | On-device |
Picovoice price vs cloud speech alternatives
A per minute bar cannot show Picovoice, because its engines run on your hardware under a quoted license. The chart instead prices the cloud route for teams willing to stream or upload audio, while the free on-device options sit in the last row.
Cloud speech to text APIs, cost per minute
Batch transcription on Hear runs $0.001 a minute
Batch list prices per minute from vendor pages and Google's published docs, checked 30 October 2026. Bar length follows price, and the last column names what each engine is chosen for.
- PyAI HearSpeech to textPhone audio tested
$0.001 a minute on Hear, with 4.76% word error on 8 kHz phone audio in PyAI tests and $50 of free credit
- AssemblyAIUniversal-3.5 ProHourly billing
$0.21 an hour batch, about $0.0035 a minute, or $0.15 an hour on Universal-2
- SpeechmaticsMelia 1Accent depth
$0.004 a minute pre-recorded, rising to $0.0067 for real time, with $100 of sign-up credit
- Deepgram Nova-3Batch and liveLive and batch
$0.0043 a minute pre-recorded, $0.0077 when streaming, with $200 of credit that never expires
- Google STT V2GCP nativeGCP billing
$0.016 a minute on standard V2 recognition per Google's published docs, sixteen times the Hear rate
- Picovoice and Vosk, WhisperOn-deviceRuns locally
No per minute meter. Picovoice licenses are quoted by sales, while Vosk and Whisper cost only the device or server you already run
Rates checked 30 October 2026. Picovoice licenses its engines per deployment, so it has no bar.
Picovoice competitors reviewed
The seven picks split by on-device and cloud. Vosk and Whisper run offline like Picovoice, PyAI, Deepgram, AssemblyAI, Google Cloud STT and Speechmatics are cloud APIs, and PyAI adds cloud phone agents.
PyAI
Built for teams moving agents and transcription to the cloud
PyAI is cloud voice infrastructure with four products on one API key: Speak for text to speech, Cast for long narration, Hear for transcription and Omni for phone agents. Against Picovoice it only covers the cloud side, so a builder who needs Porcupine style wake word on an ESP32, or audio that never leaves a Raspberry Pi, should keep Picovoice, Vosk or Whisper.
Hear batch transcription costs $0.001 per audio minute and streams a first partial in about 200 ms, with 4.76% word error on 8 kHz phone audio in PyAI tests. Omni phone agents bill $0.05/min plus $0.01 for telephony, and Speak reads text at $0.04 a minute.
PyAI launched recently and has no G2 reviews yet, so judge it on a sandbox run with your own recordings.
PyAI features
- Hear
- Batch and streaming transcription at $0.001 a minute, 92 to 247 times real time on long files.
- Speak
- Text to speech with 200 voices, so app prompts and spoken alerts come from the same key as transcription.
- Omni
- Cloud phone agents at $0.05/min + $0.01 telephony, with end of turn detected about 196 ms after the caller stops.
- OpenAI compatible
- Requests follow the OpenAI audio format, so an app already calling a Whisper style API changes its base URL rather than its code.
PyAI pricing
Hear
$0.001/min
Transcription
Speak
$0.04/min
Voiceovers
Cloning
Free
Short clip
Hear meters by the minute and there is no monthly plan. Sandbox keys never bill, and a new account opens with $50 of free credit without a credit card. Hear the voices on PyAI Speak.
PyAI pros and cons
Pros
- Hear batch transcription costs $0.001 a minute, so 10,000 minutes of recorded calls or voice notes come to $10.
- One key covers transcription, 200 Speak voices and Omni phone agents, while Picovoice has no telephony product at all.
- Usage-based billing with no monthly plan means a prototype that sits idle for a month costs nothing beyond the minutes it processed.
- Word error of 4.76% on 8 kHz phone audio and 1.59% on clean audio in PyAI tests, at parity with top cloud models.
Cons
- Every request leaves the device, so PyAI does not fit offline products, air gapped sites and privacy rules that Picovoice was picked to meet.
- No wake word engine: an always listening trigger still needs Porcupine or openWakeWord on the device before audio reaches Hear.
- A newer vendor with no G2 profile, and the 200 ms partial and 4.76% error figures come from PyAI testing, not outside benchmarks.
PyAI reviews
Nobody has reviewed PyAI on G2 yet, so this tab stays empty instead of carrying made up quotes. A free sandbox key with $50 of credit and no credit card is enough to run your own Picovoice test clips through Hear.
Choose PyAI if your audio can leave the device for the cloud and you want transcription at $0.001 a minute or a phone agent rather than an offline wake word.
Vosk
Built for free, open source on-device speech to text
Vosk is the closest open source swap for Picovoice Leopard and Cheetah. The offline recognizer runs on Android, Raspberry Pi and servers, with models near 50 MB and bindings for many programming languages.
Vosk is free under Apache 2.0, which permits commercial use with no per-unit fee, so the only real cost is the device you already ship. Vosk bundles no wake word and no text to speech, so a full local stack means pairing it with openWakeWord for the trigger and the local TTS engine for playback.
The tradeoff is accuracy and upkeep. The small models that fit a Raspberry Pi trail Leopard on noisy or strongly accented speech, and you maintain the runtime, the model files and the glue code yourself. Vosk fits a maker or an embedded team that already owns its hardware, wants audio to stay on the device, and will trade polish and support for zero license cost.
Vosk features
- Offline recognizer
- Vosk transcribes on Android, Raspberry Pi and servers with no network call, the same on device promise as Leopard and Cheetah.
- Small models
- Vosk models weigh near 50 MB, small enough for boards where larger Whisper models cannot keep up live.
- Apache 2.0 license
- No AccessKey, activation meter or sales call, so a hobby build and a 10,000 unit fleet pay the same zero in fees.
- Language bindings
- Bindings for many programming languages let Vosk drop into an existing Python, Java or C codebase.
Against Picovoice's pain points: Vosk
A free, offline drop in for Leopard and Cheetah, without wake word or voice output.
- Free tier: Vosk has no tier to expire, since the Apache 2.0 license keeps it free after the Picovoice Free Tier cutoff of 30 June 2026
- Published price: The price is zero and public, against Picovoice paid plans that begin with a sales call
- Self serve: Anyone can download Vosk and ship it today, which suits indie developers Picovoice's enterprise focus leaves behind
- On device: Vosk runs fully offline on a Raspberry Pi or Android phone, keeping audio on the device like Picovoice Leopard
- Wake word: Vosk ships no wake word, so Porcupine users need openWakeWord beside it and should test custom phrases first
- Speech output: Vosk has no text to speech, so a Picovoice Orca replacement means adding the local TTS engine or another local voice
Vosk pros and cons
Pros
- Apache 2.0 license with no activation meter, so a fleet of 10,000 devices costs the same in fees as a single prototype.
- Runs fully offline on a Raspberry Pi or Android phone, keeping audio on the device just as Picovoice Leopard does.
- Models near 50 MB fit small boards where the larger Whisper models are too heavy to keep up in real time.
Cons
- No wake word, intent or text to speech, so replacing the whole Picovoice suite means stitching openWakeWord and the local TTS engine around it.
- Accuracy trails Whisper and the cloud engines on noisy or accented speech, so test on your own recordings first.
Choose Vosk if you build on devices, need free offline speech to text, and can live without a bundled wake word.
Whisper
Built for open source transcription you run yourself
Whisper is the open transcription model from OpenAI, and the C++ port of Whisper ports it to CPU, Metal, GPU acceleration and phone NPUs. Accuracy across many languages is its strength, which makes it the pick when Leopard batch output fell short.
The weights are MIT licensed and free, so the bill is your GPU or server time. Larger models are heavy for a microcontroller and the project ships no wake word or voice output, so it suits a server or a recent phone better than an ESP32.
Whisper runs as a batch transcriber rather than a live streaming engine, so real-time captions need extra chunking work or a community fork such as whisper-streaming. Whisper fits a team that wants OpenAI-grade multilingual accuracy, controls its own GPU or server time, and can host the larger models instead of squeezing them onto a tiny chip.
Whisper features
- MIT weights
- OpenAI publishes the Whisper weights under MIT, so transcription costs only the compute you run.
- the C++ port of Whisper
- the C++ port of Whisper port builds for CPU, Metal, GPU acceleration and phone NPUs, covering laptops, phones and servers with one model family.
- Multilingual accuracy
- Accuracy across many languages is Whisper's strength, useful where Leopard batch output fell short.
- Model sizes
- Smaller Whisper models fit a recent phone, while larger ones need a GPU for real time output.
- Batch first
- Whisper is built for files, so live captions need chunking work on top.
Against Picovoice's pain points: Whisper
The free, accurate route off Leopard for servers and recent phones.
- Free tier: Whisper weights stay free under MIT, so the end of Picovoice Free Tier AccessKeys does not apply
- Published price: No license fee exists, and the real cost is GPU or server time you can measure yourself
- Self serve: the C++ port of Whisper builds without an account or contract, which suits the small device makers Picovoice no longer courts
- On device: Whisper runs locally on a phone NPU or laptop, though an ESP32 class chip is too small for the larger models
- Wake word: The Whisper project ships no wake word, so Porcupine stays unfilled unless openWakeWord is added
- Speech output: Whisper only listens, so Orca text to speech needs a separate engine
Whisper pros and cons
Pros
- MIT license and free weights, so transcription cost is only the compute you run, with no per user or per activation meter.
- the C++ port of Whisper builds for CPU, Metal, GPU acceleration and NPUs, letting one model family cover a laptop, a phone and a server.
Cons
- Larger models need a GPU for real time output, a cost Picovoice Cheetah avoids on a small chip.
- Batch first by design, so live captions need extra chunking work that a streaming API handles for you.
- No wake word, intent or speech output, so Whisper replaces Leopard but leaves Porcupine, Rhino and Orca unfilled.
Choose Whisper if you have a server or recent phone and want accurate open transcription you control.
AssemblyAI
Built for batch transcription priced by the hour
AssemblyAI sells cloud transcription by the hour, with speaker labels, summaries and other audio intelligence features layered on its Universal models.
Universal-3.5 Pro batch costs $0.21 an hour, about $0.0035 a minute, and Universal-2 costs $0.15 an hour. Streaming is billed on connection time, and sign up adds $50 of credit with no credit card. AssemblyAI offers no text to speech.
The audio intelligence layer is the real draw. Summaries, topic detection, sentiment and PII redaction arrive with the transcript, so a note-taking or call-analytics product skips a second service. The tradeoff against Picovoice is that every file leaves the device for the cloud and there is no offline or wake-word path. AssemblyAI fits a team replacing Leopard batch jobs that wants rich metadata, not just raw text.
AssemblyAI features
- Universal-2 batch
- $0.15 an hour, or $0.0025 a minute, for recorded audio archives.
- Universal-3.5 Pro
- $0.21 an hour, about $0.0035 a minute, for files that need higher accuracy.
- Streaming options
- Universal-Streaming lists at $0.15 an hour and Universal-3.6 Pro Realtime at $0.45 an hour, billed on connection time.
- Speaker labels
- Speaker labels and summaries come from the same API, so call or meeting notes need no second vendor.
- Sign up credit
- $50 of credit with no credit card covers about 14,000 minutes of Universal-3.5 Pro batch audio.
- G2 standing
- 125 G2 reviews average 4.6, with accuracy and ease of use named most often as of 15 Oct 2026.
Against Picovoice's pain points: AssemblyAI
A cloud transcription API with public rates, for teams that no longer need audio on the device.
- Free tier: $50 of credit with no credit card replaces the lost Picovoice Free Tier for testing, though the credit is one time rather than ongoing
- Published price: Universal-2 at $0.15 an hour is on the public price page, where Picovoice paid rates need a sales call
- Self serve: Sign up and an API key are enough to start, so indie teams skip the enterprise sales path Picovoice now runs
- On device: Audio goes to AssemblyAI's cloud, so AssemblyAI fits only teams whose offline rule was about cost, not privacy
- Wake word: AssemblyAI offers no wake word, so a Porcupine replacement has to come from another tool
- Speech output: AssemblyAI has no text to speech, so Orca prompts need a separate voice provider
AssemblyAI pros and cons
Pros
- Universal-2 batch at $0.15 an hour works out to $0.0025 a minute, low enough for large archives of recorded audio.
- Speaker labels and summaries come from the same API, so meeting or call notes need no second vendor.
- G2 lists 125 reviews averaging 4.6, with accuracy and ease of use named most often as of 15 Oct 2026.
- $50 of sign up credit with no credit card covers about 14,000 minutes of Universal-3.5 Pro batch audio.
Cons
- Streaming bills for connection time, not audio sent, so an idle WebSocket left open keeps costing money.
- No text to speech or wake word, so replacing Picovoice Orca or Porcupine needs separate tools.
AssemblyAI reviews
AssemblyAI holds 4.6 on G2 from 125 reviews. AssemblyAI on G2 ↗
A reviewer highlights: "It was quick for our team to integrate, and it delivers accurate transcription results even with long audio files and conversations involving multiple speakers." (Verified user, Mid-Market, G2 Apr 2026)
Another flags: "Transcription accuracy occasionally drops with poor audio quality or strong background noise." (Gaurav R., Small-Business, G2 Jun 2026)
On pricing: "Pricing is fair, and they give you a generous free trial to test your integration without having to pay." (Verified user in film, Small-Business, G2 Aug 2026)
Choose AssemblyAI if you move recorded audio to the cloud and want speaker labels and summaries with each transcript.
Deepgram
Built for Nova-3 transcription and Aura-2 voices on one account
Deepgram pairs Nova-3 transcription with Aura-2 text to speech and a voice agent API, so one cloud account can stand in for both Cheetah and Orca.
Nova-3 costs $0.0043 a minute pre-recorded and $0.0077 streaming, and Aura-2 voices cost $0.030 per 1,000 characters. New accounts get $200 of credit that does not expire.
Streaming is Deepgram's strong suit, with low-latency partials that suit live captioning and phone bots, and the non-expiring $200 credit covers a long pilot before any commitment. The catch is that Deepgram is cloud-only, so it cannot match Picovoice on fully offline or microcontroller deployments. Deepgram fits a team that wants streaming speech to text and a voice out of one vendor and is comfortable sending audio to the cloud.
Deepgram features
- Nova-3 pre-recorded
- $0.0043 a minute for recorded audio files.
- Nova-3 streaming
- $0.0077 a minute for live transcription, the closer match to Cheetah's real time role.
- Aura-2 voices
- Aura-2 text to speech at $0.030 per 1,000 characters stands in for Picovoice Orca from the same account.
- Voice agent API
- A voice agent API joins listening and speaking for assistants that once combined several Picovoice engines.
- Credit that stays
- $200 of credit that does not expire stretches to about 46,000 minutes of pre-recorded audio.
Against Picovoice's pain points: Deepgram
One cloud account that can replace both Cheetah and Orca.
- Free tier: $200 of credit that does not expire gives longer runway than most cloud trials after Picovoice Free Tier keys stopped working
- Published price: Nova-3 at $0.0043 to $0.0077 a minute and Aura-2 at $0.030 per 1,000 characters are listed publicly
- Self serve: Pay as you go needs no credit card to start, so small device teams avoid a Picovoice style sales call
- On device: Pay as you go plans send audio to the cloud, so a strict offline rule still rules Deepgram out
- Wake word: Deepgram lists no wake word engine, so Porcupine needs a local replacement such as openWakeWord
- Speech output: Aura-2 voices cover the Orca role, letting one Deepgram bill handle transcription and spoken replies
Deepgram pros and cons
Pros
- One account covers streaming transcription, Aura-2 voices and agents, replacing two Picovoice engines with one bill.
- G2 shows 482 reviews averaging 4.6, a large review base among the cloud picks here as of 15 Oct 2026.
- $200 of credit that never expires stretches to about 46,000 minutes of Nova-3 pre-recorded audio.
Cons
- Streaming at $0.0077 a minute costs nearly double the pre-recorded rate, which adds up for always on captions.
- G2 reviewers flag limited language support and pricing as their most common complaints.
- Audio goes to the cloud on pay as you go plans, so a privacy rule that ruled out cloud still rules out Deepgram.
Deepgram reviews
Deepgram holds 4.6 on G2 from 482 reviews. Deepgram on G2 ↗
A reviewer highlights: "Deepgram is really a developer-friendly solution and is extremely fast if you compare it against other vendors like OpenAI Whisper and Speechmatics." (Diego P., Enterprise, G2 Aug 2026)
Another flags: "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." (Verified user, Small-Business, G2 Aug 2026)
On pricing: "It supports word recognition and speaker diarization, and it also has a very good price offering comparatively." (Akshay M., Mid-Market, G2 Aug 2026)
Choose Deepgram if you want Cheetah and Orca replaced by one cloud bill with generous starting credit.
Google Cloud Speech
Built for speech to text inside an existing Google Cloud project
Google Cloud Speech to Text fits teams already billing through GCP, with wide language coverage and the access controls of a Google Cloud project.
Standard V2 recognition lists at $0.016 a minute per Google's published docs, with 60 free minutes a month. Setup needs a GCP project, a billing account and service account keys before the first request.
At $0.016 a minute Google STT is the priciest cloud option on this page, several times the Deepgram or AssemblyAI rate, so the pull is integration rather than cost. IAM roles, VPC controls and logging already match the rest of a Google stack. Google STT has no on-device mode, so Google STT cannot replace Leopard where audio must stay local. Google Cloud Speech fits a team standardised on GCP that values governance over the lowest per-minute price.
Google Cloud Speech features
- Standard V2
- Standard V2 recognition lists at $0.016 a minute, so 10,000 minutes cost $160 a month.
- Free minutes
- 60 free minutes a month allow small tests without a bill.
- GCP controls
- Billing, IAM and logging come from the existing Google Cloud project, which skips a new vendor review for GCP teams.
- Setup steps
- A GCP project, billing account and service account keys must exist before the first request.
Against Picovoice's pain points: Google Cloud Speech
A cloud recognizer for teams already on GCP, at a higher rate than the other cloud picks.
- Free tier: 60 free minutes a month continue indefinitely, a small but ongoing allowance compared with the ended Picovoice Free Tier
- Published price: $0.016 a minute is published, though G2 reviewers still call Google Cloud Speech expensive
- Self serve: Self serve, but a project, billing account and service account keys come first, which G2 reviewers describe as complex
- On device: Standard plans process audio in Google data centers, so Google cannot replace Picovoice on offline devices
- Wake word: Google Cloud Speech has no wake word on this page, so Porcupine needs a local substitute
- Speech output: Voice output is not covered on this card, so an Orca replacement means a second product or vendor
Google Cloud Speech pros and cons
Pros
- G2 shows 316 reviews averaging 4.5, with ease of use and real time transcription the most praised points as of 15 Oct 2026.
- Lives inside existing GCP billing, IAM and logging, so teams already on Google skip a new vendor review.
Cons
- At $0.016 a minute, 10,000 minutes cost $160 a month, sixteen times the $10 Hear bill for the same audio.
- G2 reviewers call it expensive and complex to configure, the two most repeated complaints.
- A Google Cloud project, billing account and service account must exist before the first test request.
- Standard plans process audio in Google data centers, so it cannot replace Picovoice on offline devices.
Google Cloud Speech reviews
Google Cloud Speech holds 4.5 on G2 from 316 reviews. Google Cloud Speech on G2 ↗
A reviewer highlights: "The accuracy is genuinely impressive, especially when dealing with mixed-language audio. We run a logistics platform that operates across Arabic, English, and Urdu, and the transcription quality holds up well even with regional accents and background noise. The automatic punctuation and speaker diarization save a lot of manual cleanup time." (Muhammed A., Mid-Market, G2 July 2026)
Choose Google Cloud Speech if you already run on Google Cloud and want speech to text inside the same bill.
Speechmatics
Built for accent heavy transcription on the Melia 1 model
Speechmatics built its name on accent and dialect handling, and it now sells text to speech beside its transcription models.
Melia 1 costs $0.004 a minute pre-recorded and $0.0067 in real time, and an opt-in to model training takes a third off. Text to speech is $0.011 per 1,000 characters with the first 1 million characters free, and sign up adds $100 of credit.
The accent and dialect accuracy is the reason to pick it. Products whose users speak English in many regional varieties tend to see fewer corrections than with a generic model, and the opt-in training discount lowers the bill on high-volume batch work. Like the other cloud engines here, Speechmatics sends audio off the device, so it does not cover the on-device or wake-word jobs Picovoice handles. Speechmatics fits a global consumer or enterprise product that transcribes varied accents and wants one vendor for speech to text and voice output.
Speechmatics features
- Melia 1 batch
- $0.004 a minute for pre-recorded audio, close to AssemblyAI pricing.
- Melia 1 real time
- $0.0067 a minute for live audio, about two thirds more than batch.
- Training opt in
- Opting in to model training takes 33% off the listed rates.
- Accent handling
- Accent and dialect accuracy is the reason Speechmatics gets shortlisted, backed by a 4.7 G2 average from 75 reviews.
- Text to speech
- Speech output costs $0.011 per 1,000 characters, with the first 1 million characters free.
- Free account limits
- $100 of sign up credit, with 2 concurrent real time sessions on free accounts.
Against Picovoice's pain points: Speechmatics
Cloud transcription plus text to speech with public rates and a generous voice allowance.
- Free tier: $100 of credit plus 1 million free text to speech characters gives hobby builders room to test after the Picovoice cutoff
- Published price: Melia 1 rates and text to speech pricing are on the public page, unlike Picovoice's sales led paid plans
- Self serve: Sign up is self serve, though 2 concurrent real time sessions on free accounts cap load tests of a device fleet
- On device: Audio is processed in the cloud, so Speechmatics does not cover on device recognition
- Wake word: Speechmatics offers no offline wake word, so Porcupine needs openWakeWord or a similar local engine
- Speech output: Text to speech at $0.011 per 1,000 characters can voice prompts that Picovoice Orca handled before
Speechmatics pros and cons
Pros
- Melia 1 batch at $0.004 a minute sits close to AssemblyAI while keeping the accent depth Speechmatics is known for.
- G2 lists 75 reviews averaging 4.7, the highest average among the cloud APIs here as of 15 Oct 2026.
- The first 1 million text to speech characters are free, enough to voice prompts while testing a replacement for Orca.
Cons
- Free accounts allow only 2 concurrent real time sessions, which caps load tests of a live device fleet.
- Real time Melia 1 at $0.0067 a minute costs about two thirds more than the same model in batch.
- G2 reviewers mention limited features and slow performance among the cons.
- Audio is processed in the cloud, so Speechmatics covers neither offline wake word nor on-device recognition.
Speechmatics reviews
Speechmatics holds 4.7 on G2 from 75 reviews. Speechmatics on G2 ↗
A reviewer highlights: "We appreciate the relative ease of integrating Speechmatics into our broadcast monitoring workflow. Customer support is reliable and provides quick responses." (Joe T., Small-Business, G2 Aug 2025)
Another flags: "It appears to fill in a whole lot of nonsense when it hears noise or background chatter, creating disconcerting false positive transcriptions." (Verified user, Mid-Market, G2 Sep 2026)
On pricing: "Functionalities and pricing tailored to your exact application and business domain, a good software stack and API, reliable transcoding." (Verified user in broadcast media, Enterprise, G2 Feb 2023)
Choose Speechmatics if your users speak with strong accents and you want batch transcription near $0.004 a minute.
Picovoice alternatives by job and use case
Match the Picovoice engine you are replacing to the tool that does that one job well.
Which tool replaces each Picovoice job
| Job | Pick | Why |
|---|---|---|
| Batch transcription in the cloud | PyAI Hear | $0.001 a minute batch, 10,000 minutes for $10 |
| Speaker labels and summaries | AssemblyAI | $0.21 an hour on Universal-3.5 Pro |
| Offline speech to text, the Leopard swap | Vosk, Whisper | Free, audio never leaves the device |
| Phone agents with telephony | PyAI Omni | $0.05/min + $0.01 telephony |
| Strong accents and dialects | Speechmatics | Melia 1 from $0.004 a minute |
| Always on wake word, the Porcupine swap | openWakeWord | Code is Apache 2.0, but the pretrained wake-word models are non-commercial (CC BY-NC-SA 4.0), so train your own to ship commercially |
Speech to text by price
| Tool | Price per minute |
|---|---|
| PyAI Hear | $0.001, billed per minute |
| AssemblyAI | About $0.0035 on Universal-3.5 Pro |
| Speechmatics | $0.004 on Melia 1 |
| Deepgram Nova-3 | $0.0043 pre-recorded |
| Google V2 | $0.016 on standard recognition |
Open source on-device voice stack
| Model | License | Note |
|---|---|---|
| openWakeWord | Code Apache 2.0 | Wake word, the Porcupine swap. Pretrained models are non-commercial CC BY-NC-SA 4.0, so train your own to ship commercially |
| Vosk | Apache 2.0 | Offline STT with models near 50 MB |
| the C++ port of Whisper | MIT | Whisper on CPU, Metal, GPU acceleration and NPUs |
| the local TTS engine | GPL-3.0 | Local text to speech in place of Orca (active OHF-Voice/piper1-gpl build. The older rhasspy/the local TTS engine was MIT) |
Free and open source Picovoice alternatives
The usual 2026 replacement for a lapsed Picovoice free key is openWakeWord for the trigger, Vosk or the C++ port of Whisper for transcription and the local TTS engine for speech, all self hosted on hardware you already own. Cloud trials cost nothing up front either: PyAI gives $50 with no credit card, Deepgram $200, Speechmatics $100, AssemblyAI $50 and Google Cloud 60 free minutes a month.
When Picovoice is still the right pick
Picovoice is still the better choice when audio must stay on the device, as in wearables, cars, medical devices and offline kiosks, where one vendor covers wake word, speech to text, intent and text to speech in a single SDK. Home Assistant users also report Porcupine catching custom wake words more reliably than openWakeWord. Teams with no offline rule gain the most from a per minute cloud API.
How to move from Picovoice in 4 steps
Plan about a week for a device fleet, since wake word models need retesting on real hardware.
Step 1 of 4
List your engines
Note which Picovoice engines you call, such as Porcupine, Leopard or Orca, and whether each one has to keep running offline.
Step 2 of 4
Split device from cloud
Keep wake word on the device with openWakeWord or Porcupine, and mark the transcription and speech jobs that can move to an API.
Step 3 of 4
Replay your recordings
Run the same 50 recorded clips through Vosk and a cloud API, then compare word error, latency and cost per minute.
Step 4 of 4
Move one product line
Ship the new stack on one device model or app first, and retire the Picovoice key once accuracy holds for a full release.
Which Picovoice alternative to pick
Choose by where the audio may go. Vosk, the C++ port of Whisper and openWakeWord keep it on the device for free. If the cloud is allowed, AssemblyAI and Deepgram are strong transcription APIs, and PyAI Hear handles batch audio at $0.001 a minute with Omni phone agents at $0.05/min + $0.01 telephony. A PyAI sandbox key costs nothing and opens with $50 of credit.
More text to speech comparisons
Picovoice alternatives FAQs
Common questions from teams comparing Picovoice.
What is a good Picovoice alternative?
Vosk or Whisper for offline, a cloud API otherwise. Vosk and the C++ port of Whisper run on the device like Picovoice and cost nothing to license. When audio may leave the device, AssemblyAI and Deepgram are proven transcription APIs, and PyAI Hear transcribes batch audio at $0.001 a minute. Picovoice itself still leads for embedded wake word.
Is PyAI an on-device alternative to Picovoice?
No, PyAI runs in the cloud. Picovoice processes audio offline on the device, and PyAI cannot do that. PyAI suits teams allowed to send audio to an API who want transcription at $0.001 a minute or phone agents at $0.05/min + $0.01 telephony. Offline products should stay with Picovoice, Vosk or Whisper.
Does Picovoice still have a free tier?
No, the Free Tier ended in 2026. Picovoice Free Tier AccessKeys stopped working after 30 June 2026, according to Home Assistant community threads quoting the company, and no non-commercial tier is planned. Paid plans are quoted through sales, so hobby projects usually move to openWakeWord, Vosk or a cloud API with free credit.
What is a free, open source Picovoice alternative?
openWakeWord, Vosk, the C++ port of Whisper and the local TTS engine. openWakeWord replaces Porcupine for wake word, Vosk handles light offline speech to text, the C++ port of Whisper brings higher accuracy on stronger hardware, and the local TTS engine speaks the replies. Together they rebuild most of a Picovoice pipeline for free, paid for only in your own device or server time.
Which cloud API costs less for transcription?
Rates run from $0.001 to $0.016 a minute. Batch list prices checked on 30 October 2026: PyAI Hear $0.001 a minute, AssemblyAI Universal-3.5 Pro $0.0035, Speechmatics Melia 1 $0.004, Deepgram Nova-3 $0.0043 pre-recorded and Google V2 $0.016. All five process audio in the cloud, so none of them replaces an offline Picovoice build.
Does Picovoice do wake word and voice agents?
Yes, but only on the device. Picovoice offers Porcupine wake word and Rhino intent offline, not a phone agent. For calls over telephony, PyAI Omni, Retell AI and Vapi run cloud agents. Cloud APIs do not listen for wake words, so Porcupine or openWakeWord stays the trigger on always listening 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.