Loading today’s stories
Build1 distinct publisher3 min readUpdated
Price, streaming transport and HIPAA paperwork do not convert into each other, so a shortlist is an intersection. Follow it through and the compliant socket costs 7.5 times the floor rate.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Run the filters against one requirement set and the survivor list gets short in a way no ranked list can show. Say the job is a live phone agent, the prompts carry SSML because somebody has to spell out an account number, and legal wants a signed BAA on file rather than a blog post.
Google Cloud Text-to-Speech sits at the bottom of the price range, holds a socket open, and has no stated concurrency ceiling [19], and the product itself is named on Google's HIPAA BAA covered-products list [11]. It still fails, because SSML is not supported on streaming requests to Chirp 3: HD [9]. You get the markup or you get the stream.
Amazon Polly matches that price and appears on the AWS HIPAA-eligible services list [11], and the reason it fails is worth stating carefully, because the usual description of Polly is wrong. There is no WebSocket TTS API, but StartSpeechSynthesisStream is a genuine bidirectional stream over HTTP/2: text goes in as events, audio comes back as it becomes available [8]. The transport is adequate. The ceiling is not, at generative engine only, 8 transactions per second and 8 concurrent requests [8]. Eight simultaneous callers is a pilot.
Keep walking the table and the first row that gives you a socket, no published ceiling and a signed BAA is Rime Mist v3 at $30 per million characters [20], which is 7.5 times the floor rate [4]. That multiple belongs to the requirement, not to the vendor, which is exactly why an ordered list cannot express it.
The middle transport is where teams talk past each other. OpenAI streams a chunked response body, and its own documentation recommends wav or pcm output because those start playing sooner than a compressed container [6]. That is meaningfully better than waiting for a finished file and meaningfully worse than a held-open socket [7], and the phrase "supports streaming" covers both without distinguishing them. OpenAI will also sign a BAA for the API without an enterprise agreement [12], which is less common than a page full of compliance logos implies. Rime is the one entry publishing dates instead of badges: compliant since February 2024, most recent audit March 2026, BAA on the Enterprise plan, with VPC or full on-prem [13]. Dates can be checked against reality.
One caution on the whole table. It is a single publisher's reading of pricing and compliance pages as published at the time of writing [22], and the quoted span is slightly off: ElevenLabs at $166.11 against a $4 floor is 41.5x, not 40x [1]. That is a trivial discrepancy, and it is also the signature of a table assembled by hand from vendor pages, which is the only way this table can be assembled. Re-derive the three columns for your own requirement before anyone signs, because the column that eliminates your cheapest option is usually the one nobody priced.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
dev.to argues that the three things deciding a TTS vendor are measured in units that do not convert: price in dollars per million characters, transport as a shape (held-open socket, chunked body, finished file), and compliance as a document that either exists or does not. With no exchange rate there is no ordering, so a ranked list must pick one axis and treat the others as tiebreakers, when in fact they are filters that compose by intersection.
Price spans about 40x: Google Cloud's legacy voices and Amazon Polly's standard engine sit at $4 per million characters, while ElevenLabs is $166.11 at its Scale tier.
Transport comes in three shapes and the difference is architectural, not incremental: WebSocket streaming holds a connection open and pushes audio as it is synthesised, chunked REST streams the response body back progressively, and batch returns a finished file.
OpenAI works by chunked REST, and its docs recommend wav or pcm output specifically because those start playing sooner than a compressed container.
Chunked REST is meaningfully better than waiting for a whole file and meaningfully worse than a held-open socket.
Amazon Polly has no WebSocket TTS API, but StartSpeechSynthesisStream is a real bidirectional streaming API over HTTP/2 in which text is sent incrementally as events and audio is received as it becomes available; it will serve a conversational agent. The constraint is the ceiling: generative engine only, 8 transactions per second, up to 8 concurrent requests.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific and checkable, but entirely single-source
The piece is unusually concrete for a vendor-comparison post: named per-million-character rates for eleven providers, named streaming mechanisms (StartSpeechSynthesisStream over HTTP/2, Chirp 3: HD), named quantitative ceilings (8 TPS, 8 concurrent requests), dated compliance attestations, and a self-contained script that reproduces the intersection result. Every figure is nonetheless the author's reading of vendor documentation, restated without links, quotes or independent corroboration from any second publisher; the arithmetic that can be checked internally holds (41.5x span, $16,200 gap, 7.5x Rime premium). The evidence is verifiable in principle rather than verified here, and it is explicitly a time-of-writing snapshot.
No adoption evidence supplied
The cluster contains no deployment, usage, customer-count, traffic or revenue evidence for any provider named. Published price lists, streaming API availability and HIPAA attestations describe vendor capability and packaging, not uptake, and the intersection result is an illustration over a hand-built catalogue rather than an observation of what buyers actually chose. No adoption observations could be recorded without inventing facts.
Broadly aligned, if anything mildly understated
The framing is deflationary rather than promotional: it attacks ranked 'best TTS API' listicles, corrects two commonly overstated transport claims in the direction of less drama, hedges every price and compliance position as published at the time of writing, and rounds its own 41.5x price spread down to 'about 40x'. The one place claims run slightly ahead of evidence is presentational: a single-row intersection produced by four author-chosen constraint values is printed as 'the real output', with no sensitivity analysis and no quality or latency axis, so a reader could over-read Google Cloud TTS as a general answer rather than the answer to one constraint set. Net position is close to aligned, tilted marginally to understatement.
Undisclosed but visible editorial tilt
No commercial relationship is disclosed anywhere in the post, and the analytic frame is genuinely vendor-neutral in construction (the constraints are user-supplied and the winning row is a hyperscaler). Two observable features still create moderate incentive risk: the catalogue is hand-assembled by the author, who therefore controls which axes exist and which vendors appear, and Rime receives distinctly promotional treatment — 'the strongest specialist position', two catalogue rows, and the dek's headline framing of the compliant socket premium. Published on a developer content platform with no editorial verification layer, that is enough to warrant scepticism, not enough to establish a sponsored piece.
Moderate — sound method, thin sourcing
Confidence is held down by structural sourcing limits rather than by internal weakness. One publisher, one article, no vendor confirmation, no second reading of any price or limit, and no adoption evidence at all; the pricing and compliance layer is a dated snapshot in a market that reprices frequently. Against that, the reasoning is transparent and reproducible, the arithmetic checks out, and the most load-bearing technical claims are specific enough to be falsified quickly by anyone with the vendor docs open. The method is more trustworthy here than the numbers.
build
The demo-best voice engine finished last: 12,247 calls argue for buying on completion rate1 distinct publisher
build
48 startups, 4 known by name, 28 recommended by category1 distinct publisher
build
Voicebot amnesia is a telephony bug: FreeSWITCH's ESL socket, not the model1 distinct publisher
build
S3 annotations move the label without moving the bytes, and checksums cannot see it1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 21, 2026