TTS API concurrency and rate limits, October 2026: which plans run 20 concurrent streams

On 1 October 2026, SpeechifyAI’s $99 Pro plan allowed 30 concurrent text-to-speech requests, streaming included, where ElevenLabs’ $99 Pro plan allowed 20 on its Flash models and no cheaper subscription plan from any provider reached 20. SpeechifyAI’s $10 Starter plan allowed 15, and no other provider’s subscription plan under $99 a month allowed more than 10 (snapshot).

The cheapest routes to 20 concurrent streams carry no plan fee at all. Rime’s usage-billed Starter plan allows 20, Amazon Polly’s generative voices 26 over SynthesizeSpeech (its bidirectional streaming API allows 8), Deepgram’s pay-as-you-go streaming API 45 and Google Cloud’s Chirp 3 HD streaming 100 sessions per project (snapshot). Among monthly subscriptions, $99 is the floor: SpeechifyAI’s Pro plan allows 30 and ElevenLabs’ Pro plan 20 on Flash, while Inworld’s $100 Builder plan allows 50.

At 15 concurrent requests, the cheapest subscription after our $10 Starter plan was ElevenLabs’ $99 Pro plan, on its Flash models. Cartesia needs its $299 Scale plan for 15, and Hume its $200 Scale plan, which allows 20 (snapshot).

Below: every plan’s limit side by side, the request-rate limits, how each provider counts a request, what happens past the limit, the availability commitments each one publishes, and where SpeechifyAI is behind.

Every plan’s limit

Each self-serve plan’s monthly price and the concurrent text-to-speech requests it allows, as each provider published them on 1 October 2026 (snapshot).

Concurrent requests are not concurrent calls

A voice agent synthesizes speech only while it is talking, so one concurrent request can carry several calls. Cartesia puts the parallel conversations a plan supports at “effectively 4X your concurrency limit”, and Inworld puts concurrent user sessions at “typically at least 4x your concurrent request limit” (snapshot). That ratio is arithmetic, not a feature, and it holds for every API in this post, ours included, because it depends on how much of each call is spent speaking. Microsoft’s quota guide works the same arithmetic for a call center: 1,000 concurrent calls, with agents speaking half the time in five-second replies, need 100 text-to-speech transactions a second. On the SpeechifyAI API, a request holds its slot until its response has finished, so a long `/v1/audio/stream` response holds one for the whole stream.

Request-rate limits

A concurrency limit caps how many requests run at once, and a rate limit caps how many start each second or minute. Azure, OpenAI and MiniMax publish only a rate limit, and Google, Amazon and Hume publish both (snapshot).

How each provider counts a request

Two plans with the same number can behave differently, because providers count different things (snapshot):

• Per account, across endpoints: SpeechifyAI counts every synthesis endpoint, streaming included, against one budget per account, and Inworld’s limits “apply per account and are shared across your API keys”.
• Per model family: ElevenLabs sets one limit for its Flash models and a lower one for Multilingual v2, and states no separate figure for Eleven v3.
• Per connection type: Deepgram allows 15 concurrent REST requests on its self-serve plans and 45 or more streaming ones, and Cartesia counts the contexts active at once, where every HTTP request is its own context.
• Per project, or per account and region: Google’s quotas apply to each Cloud project and Amazon Polly’s to each account in each region, and both are raised through a console request.
• By total spend: Fish Audio’s tiers unlock “as soon as your total prepaid amount reaches the threshold”.

What happens past the limit

Past the limit a provider either refuses the request or holds it, and the two fail differently under load (snapshot):

• Refused with 429: SpeechifyAI, Cartesia and Inworld return 429 Too Many Requests, and ours carries a Retry-After header and an error code, rate_limited or concurrency_limit_reached, that says which limit was hit.
• Queued: ElevenLabs processes requests past its limit “in a queue alongside lower-priority requests”, which it says “typically only adds ~50ms of latency”.
• Throttled with HTTP 400: Amazon Polly returns a ThrottlingException.
• Refused for capacity, not quota: Microsoft says most 429 errors on Azure’s standard voices come from “limited backend service capacity for a specific voice in the selected region”, which a higher quota does not cure.
• Not stated: The pages we read from Rime, Deepgram, Hume, Gradium, Fish Audio and Google do not say.

A refusal shows up in a load test as errors and a queue as latency, so test at the concurrency you expect and watch both.

Published availability commitments

On the pages we read on 1 October 2026, the providers’ availability commitments were these (snapshot):

• Amazon Polly is covered by AWS’s Machine Learning Language SLA, which credits 10% of a month’s charges when a region’s monthly uptime falls below 99.9%, 25% below 99.0% and 100% below 95.0%.
• Deepgram lists “Standard Uptime” on its self-serve plans, without a percentage.
• ElevenLabs, Inworld and Rime offer SLAs on Enterprise contracts.
• Cartesia, Hume, Gradium and Fish Audio show no SLA on the pricing and limits pages we read.
• Google and Microsoft publish SLA documents for their speech services, but neither rendered for our reader, so their terms are not quoted here.
• SpeechifyAI publishes neither an uptime percentage nor an SLA, and our pricing page lists dedicated support on the Scale plan.

Where SpeechifyAI is behind

The same reading shows where we are behind (snapshot):

• The free plan: Ours allows 3 concurrent requests and 1 request a second, and pauses at 500,000 characters a month, while Rime’s Starter allows 20 and Deepgram’s pay-as-you-go streaming API 45, billed per use with no plan fee.
• Around $100 a month: Inworld’s $100 Builder plan allows 50 concurrent requests, against 30 on our $99 Pro plan.
• Ceilings raised on request: Azure raises its Standard text-to-speech rate up to 1,000 transactions a second on request, and Google and Amazon raise their quotas through a console, while our limits rise with the plan and Enterprise limits are set in the contract.
• No uptime commitment: AWS publishes the 99.9% line where its Polly credits start, and we publish neither an uptime percentage nor an SLA.

SpeechifyAI’s limits by plan

Our limits are counted per account and shared by every text-to-speech endpoint: /v1/audio/speech, /v1/audio/stream and /v1/audio/stream/with-timestamps (API limits).

Every response carries RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset, so a client can pace itself before it is refused. Upgrading in the console raises the limits immediately. Inside those limits, we can also set a limit for each end user of your own app, on request.

How we read the limits

We read each provider’s own documentation and pricing pages on 1 October 2026 and kept every figure, with a verbatim quote and its source page, in the snapshot. Every concurrency figure was read at least twice, and where two reads disagreed, the snapshot records which one we kept and why. Prices are the monthly-billed price in US dollars, before tax. Pages our reader could not open are listed in the snapshot, and nothing from them is used: ElevenLabs’ help center article on request limits, Google’s SLA and pricing pages, Microsoft’s SLA document and OpenAI’s Scale Tier page. Enterprise contracts, regional differences and quotas raised on request are out of scope. What these APIs cost at 1M, 10M and 100M characters a month is in our TTS API pricing comparison. How text-to-speech models score with blind listeners is in our comparison of TTS APIs by Elo and price.

Size for your busiest minute

Count the streams your product holds open in its busiest minute, add headroom, and choose the plan that covers them, knowing whether its provider refuses or queues the rest. You can start free with 3 concurrent requests and 500,000 characters a month, then move to Starter for 15 or Pro for 30 when your traffic needs them.