best 9 speech-to-text apis

gladia states that it does not train on your audio, that this is the default rather than an upgrade, and that it is verifiable in the data processing agreement. every rival either trains by default with an opt-out you have to find, or declines to say. if the recordings belong to your customers rather than to you, that difference is the product.

diarization is included at every tier at no extra charge, where deepgram charges $0.12 an hour for it and assemblyai up to $0.12 on top of streaming. 100+ languages with mid-sentence language switching, and a self-reported 9.6% word error rate on real english audio for its solaria-3 model.

the price is the trade. $0.61 an hour pay-as-you-go is the highest entry rate in this ranking — six times rev ai's cheapest tier — and the $0.20 figure on the marketing page needs a committed volume contract. read the headline as the committed price, not yours.

$0.15 an hour on universal-2 and $0.21 on the newer universal-3.5 pro are the cheapest fully-managed batch rates among the pure-play vendors, and assemblyai is refreshingly the only one that publishes in dollars per hour rather than making you convert. 99+ languages on streaming.

the streaming gotcha is worth a paragraph of anyone's attention: real-time is billed by how long the websocket stays open, not by how much audio you send. an hour-long connection carrying thirty minutes of speech bills a full hour. for a call-centre application with hold music and silence, the effective rate is well above the quoted $0.45.

on data, assemblyai's terms grant it a licence to use customer data including for model training, with an opt-out available through account settings or support. that is the common position in this category — better than deepgram's price-linked version, worse than gladia's default.

scribe v2 ranks first for word error rate on artificial analysis's aa-agenttalk dataset at 1.5% — an independent leaderboard result rather than an elevenlabs claim, which is rare enough in this category to be the headline. diarization handles up to 32 speakers with per-word and per-character timestamps, plus audio-event tagging for things like laughter and music.

the billing trap is severe and easy to walk into. pay-as-you-go is $0.22 an hour. paying with the credits bundled into a subscription costs about $1.98 an hour, because scribe burns 330 credits per minute and those credit pools are sized for text-to-speech character volume rather than transcription minutes. same model, same output, nine times the cost, and the plan page won't tell you.

no self-hosting, and elevenlabs' training policy for scribe audio specifically wasn't confirmable on its pages — a gap worth closing yourself if the recordings are sensitive.

whisper's weights are mit-licensed across every size from tiny to large-v3, so the model openai bills $0.36 an hour for is one you can run on your own gpu for the cost of electricity. no other vendor here offers that exit, and it means every managed price on this page should be read as a convenience premium over a known free alternative.

openai's api data controls state that audio sent to the transcription endpoints is not used for training unless you opt in, with roughly 30-day retention for abuse monitoring and stricter terms available. that is the correct default and only gladia matches it.

two real gaps. there is no native speaker diarization at all — you bolt on a separate tool — and there's no free tier or trial credit for the transcription endpoints. worth knowing too that whisper large-v3 is no longer state of the art: the hugging face open asr leaderboard now places cohere transcribe and ibm granite speech above it, though whisper remains the most widely deployed baseline.

$0.10 an hour on reverb turbo is the lowest published rate anywhere in this ranking, with the standard reverb model at $0.20 and whisper-based options at $0.30. free credits equivalent to five hours of reverb make it cheap to evaluate.

rev describes reverb and reverb turbo as open-source models, which would give it a self-hosting path alongside whisper's — but we could not confirm the licence or repository within budget, so treat that as a promising lead rather than a fact. its marketing cites gains 'over competitors' rather than an absolute error rate against a named dataset, so there's no accuracy figure here we'd publish.

the pricing page is built around batch. no separate real-time rate appears despite a documented streaming api, and while diarization is marketed as part of the model, no line item confirms whether it's free or billed. 57+ languages async but only 9+ on streaming, which is a sharp drop.

the streaming rate of $0.288 an hour is the lowest real-time figure in this ranking, the pricing table itemises add-ons cleanly, and $200 of free credit makes evaluation easy. on presentation alone deepgram looks like the most transparent vendor here.

then you find the condition. standard rates assume enrolment in deepgram's model improvement partnership programme — your audio helps train their models. you can opt out by passing mip_opt_out=true, and doing so forfeits the roughly 50% discount that enrolment buys. so the real choice is $0.288 an hour with training or about double without, and only the first number appears in the comparison tables everyone publishes.

diarization is a further $0.12 an hour, where gladia includes it. on-premise deployment exists but pricing is entirely sales-gated, and nova-3 multilingual covers around 30 languages, fewer than most rivals here.

the deployment options are the reason to look: saas, on-premises, and on-device, aimed squarely at buyers who cannot send audio to a third party. $100 of free credit with no card is generous for evaluation, billing is per second, and volume discounts apply automatically above 500 hours a month.

the pricing page states 'from $0.129/hr' on the pro tier and then stops. it doesn't separate batch from real-time — the only vendor here that doesn't — and a $0.24 real-time figure we found elsewhere on the site couldn't be confirmed against a fetchable page. the accuracy documentation returned 404.

the training default is likewise unresolved. data logging is described as a toggle customers can change at any time, but whether it starts on or off wasn't confirmable, which for a vendor selling to regulated buyers is the one thing that should be stated plainly.

the disconnected containers are a real differentiator. azure will let you run speech recognition fully offline in an air-gapped environment, subject to microsoft approval and a commitment plan, which almost no pure-play vendor offers. connected containers need no special approval. for defence, health and government work that can settle the decision on its own.

the pricing is unreadable. every figure on the calculator is a '$-' placeholder until you pick a region and currency, and across three attempts nothing static could be extracted. the numbers that do circulate through microsoft's own community and support channels disagree with each other by a factor of two — about $0.18 an hour against about $0.36 for standard batch — which suggests an unannounced change nobody has documented.

what we could confirm: the free f0 tier gives five audio hours a month of real-time only, batch isn't supported on it, and batch diarization is included at no extra charge on standard and custom tiers while real-time diarization is billed as an enhanced add-on.

the chirp 2 and chirp 3 models are competitive and the gcp integration is the deepest here for anyone already on google's stack. there's also a distinct on-premise product line, which most pure-play vendors don't offer at all.

we could not extract a single price. the pricing page is calculator-driven and returned no static figures across three attempts, including following the v2 documentation redirect. the widely-quoted 60 free minutes a month couldn't be confirmed on google's own page either, and google's docs point you to an internal model-comparison tool rather than publishing a headline accuracy figure.

one confirmed improvement worth noting: billing granularity moved from 15-second to 1-second rounding, which materially helps anyone transcribing lots of short clips. diarization is an available request parameter, though whether it carries a surcharge wasn't confirmable.