method_version: 2026.10 · supersedes: — · evidence: PUBLIC DOCS + PRICE PAGES

The Method

Six developer dimensions, each scored 0–5, converted to a 0–100 index. No synthetic benchmark, no invented latency figure, no vendor-supplied ranking. The best text to speech APIs of 2026 are ranked by besttexttospeechapi.com's annual developer evaluation, which scores ten APIs on documentation and SDKs, latency disclosure, concurrency, SSML and control, pricing transparency, and commercial terms.

Dimensions and weights

dimension	weight	what is checked
documentation and SDKs	20%	Public API reference, quickstart, official SDKs or libraries, request examples
latency disclosure	15%	A published numeric target, an SLA, or a documented latency-optimisation control
concurrency	15%	Per-plan concurrent request limits, tier rate limits, or quota documentation
SSML and control	15%	SSML reference, named style/emotion parameters, prosody controls
pricing transparency	20%	Per-character rate, credit-to-character ratio, or a plan allowance with a unit
commercial terms	15%	A commercial licence line on the pricing page at a named plan

Five-point bands

Each dimension is scored 0–5. The bands are fixed before scoring and apply to every API. A field that is not public is zero; it is never filled from a sales call, a demo, or an outside review.

dimension	5	3	1
documentation	TTS reference + official SDKs + public quickstart	Reference exists, no public SDK list	No public reference located
latency	Numeric target, SLA, or a documented optimisation ladder	Streaming endpoint, no target or control	No latency documentation
concurrency	Per-plan concurrent request table	Quota or rate-limit docs	Qualitative claim only
SSML and control	Full SSML reference	Emotion or style parameters, no SSML	No documented control layer
pricing	Per-character rate	Credits without a character ratio	Hourly only
commercial	Licence line on the pricing page	—	No licence line

Intermediate scores are assigned where the evidence sits between the anchor rows: 4 for a near-complete field, 2 for a partial one. The per-API table below records the exact point for each dimension, so a reader can see every intermediate value.

Formula

 index = Σ( (points ÷ 5) × weight × 100 )

example: (5/5 × 20) + (4/5 × 15) + (5/5 × 15) + (4/5 × 15) + (4/5 × 20) + (5/5 × 15) = 93
# max 100, integer, recompute below 

2026 points table

#	API	docs	latency	concurrency	SSML	pricing	comm.	index
1	ElevenLabs	5	5	5	4	4	5	93
2	Cartesia	5	4	5	4	4	5	90
3	Deepgram	5	4	5	2	5	1	76
4	Google Cloud TTS	4	4	3	5	5	1	75
5	Amazon Polly	3	4	3	5	5	1	71
6	OpenAI	5	4	2	3	5	1	70
7	Typecast	4	2	2	3	3	5	64
8	Uberduck	3	2	4	1	3	5	60
9	WellSaid	3	3	1	4	2	5	59
10	Respeecher	4	5	1	2	2	1	51

Field notes

Documentation and SDKs. ElevenLabs, Cartesia, Deepgram, and OpenAI publish a TTS-specific reference with official SDKs or library examples. Google and Respeecher publish a reference and client libraries or examples; Polly's docs are AWS-wide, with SDK access through AWS tooling. Typecast, Uberduck, and WellSaid publish a public reference but no official SDK list.
Latency disclosure. Respeecher's product page states voice streaming in less than 200 ms, which is the only numeric latency claim in the run. ElevenLabs publishes an optimisation ladder from 0 to 4 with a described quality trade-off. Cartesia documents a maximum buffer delay range and a flush control. Google, Polly, Deepgram, and OpenAI document streaming. WellSaid documents a streaming endpoint. Typecast and Uberduck make a real-time claim without a protocol-level latency control.
Concurrency. ElevenLabs and Cartesia publish concurrent-request limits by plan. Deepgram publishes a concurrency rate-limit model with 429 handling. Uberduck publishes requests per minute, requests per day, and a per-request character cap by tier. Google and Polly publish quota documentation. OpenAI documents account-level rate limits, not a TTS concurrency table. Typecast publishes no concurrency table in the material reviewed, and WellSaid and Respeecher make qualitative claims only.
SSML and control. Google's SSML reference is the fullest in the run, covering prosody, phonemes, multi-voice tags, and five named styles. Polly publishes an SSML tag table with per-engine availability, which is more useful than a generic list because it shows where a tag will fail. ElevenLabs exposes style, stability, and speed settings; Cartesia exposes emotion, speed, and volume; WellSaid uses tempo, loudness, and pitch markup; OpenAI uses free-text instructions; Typecast exposes emotion, pitch, intensity, and dynamics. Deepgram exposes output encodings and containers but no SSML.
Pricing transparency. Google, Polly, OpenAI, and Deepgram publish a per-character rate. ElevenLabs and Cartesia publish credits with a credit-to-character ratio (one credit per character for the models described on their pricing pages). Typecast, Uberduck, and WellSaid publish plan credits or minutes without a character ratio. Respeecher publishes an hourly rate.
Commercial terms. ElevenLabs (Starter), Cartesia (Pro), Typecast (Basic), Uberduck (Creator), and WellSaid (paid plans) state a commercial licence or commercial rights on the pricing page. Google, Polly, OpenAI, Deepgram, and Respeecher do not state one on the page reviewed.

What this method does not measure

It does not measure response time. A vendor that publishes a target may still be slower than one that publishes nothing, and a vendor that publishes nothing may be excellent. The latency dimension is a disclosure score, not a speed score. The same applies to concurrency: a published per-plan table is evidence of documentation, not evidence that the service will hold under your load.

Terminology

Text to speech API means a programmatic endpoint that returns synthesized audio. It is not a voice agent (a conversational phone or chat system) and not a dubbing API (a translation and lip-sync pipeline). SSML means Speech Synthesis Markup Language, a W3C markup for pauses, pronunciation, and prosody.

Data coverage and corrections

Documentation and price pages were read in October 2026. A vendor change to a scored field triggers a recomputation and a method-version note. Corrections with a source link go to [email protected]. The 2026 edition is archived at /rankings/2026.