DEV Community: unicodefawn

Free speech-to-text and text-to-speech API quotas (Oct 2026, with the catches)

unicodefawn — Mon, 12 Oct 2026 08:15:58 +0000

I needed voice input and spoken replies for a small side project, and I didn't want to pay before knowing if anyone would use it. So I went through the pricing pages of the usual speech APIs and wrote down what you actually get for free. All numbers below are rounded and were checked on Oct 12, 2026. They change often, so treat this as a starting point and confirm on the provider's page before you build on it.

Free speech quotas come in four shapes, and the shape matters more than the headline number:

A rate-limited free API that resets every day (Groq).
Free minutes or characters every month on a big cloud, usually with billing enabled (Google Cloud, Azure).
A one-time credit that lasts until you spend it (Deepgram, AssemblyAI).
A shared monthly credit pool across several products (ElevenLabs).

Groq: Whisper with daily limits

Groq's rate-limit docs list whisper-large-v3 and whisper-large-v3-turbo at 20 requests per minute, 2,000 requests per day, 7,200 audio seconds per hour and 28,800 audio seconds per day. That is about 2 hours of audio per hour and 8 hours per day. Limits apply per organization, not per user, and the docs say your exact numbers are on the account's limits page. When you go over, you get a 429 with a retry-after header, so build a retry from day one.

This is the best free option I found for prototypes that transcribe short clips all day long.

Google Cloud: monthly free tier, billing required

Speech-to-Text: the V1 API gives the first 60 minutes per month free, then $0.016 per minute with data logging or $0.024 without. The V2 price table has no free row, so if you start on V2 you pay from the first minute.

Text-to-Speech is more generous. Standard and WaveNet voices include 4 million characters per month, Neural2 and Chirp 3 HD include 1 million. The newer Gemini-TTS models have no free usage at all. Google states you must enable billing to use Text-to-Speech, and you get charged automatically past the free amount. Set a budget alert before you test anything in a loop.

Azure Speech: the F0 tier

Azure's free F0 tier includes 5 audio hours per month of real-time speech-to-text (shared between standard and custom models, batch transcription not supported) and 0.5 million characters per month of neural text-to-speech. Speech translation gets its own 5 hours. If your app needs batch jobs, F0 won't cover it.

Deepgram and AssemblyAI: one-time credits

Deepgram gives new accounts $200 in credit with no card required, and says it does not expire. Their own estimate is roughly 43,000 minutes (700+ hours) of Nova transcription. Promotional credits can't be moved between accounts.

AssemblyAI gives $50 in free credits, no card. At the listed $0.15 per hour for Universal-2, that is about 330 hours of pre-recorded audio. At $0.21 per hour for Universal-3.5 Pro, about 240 hours. Two catches from their docs: multichannel audio is billed per channel, so a stereo file costs double, and streaming is billed by how long the WebSocket stays open, not by how much audio you send. Free accounts can also open only 5 new streams per minute.

ElevenLabs: one pool for everything

The Free plan has 10,000 credits per month, shared across text-to-speech, speech-to-text, music and sound effects. Text-to-speech costs about 1 credit per character, which ElevenLabs puts at roughly 10 minutes of speech. Speech-to-text costs 330 credits per minute, so the whole pool would buy about 30 minutes of transcription. Free credits don't roll over, and the commercial license starts on the paid Starter plan. Good for trying voices, not for shipping.

The no-quota option: run Whisper yourself

OpenAI's Whisper is open source, and whisper.cpp runs it on a normal laptop CPU. There is no monthly cap, no API key and no audio leaving your machine. The cost is your hardware and speed. For batch jobs like transcribing old recordings overnight it is hard to beat, and it pairs well with a hosted free tier for the real-time part.

What I ended up using

Groq for live transcription in the prototype, since the daily reset fits short clips. Google's Standard voices for spoken replies, with a budget alert set to a few dollars. Local Whisper for the backlog of existing audio. If the project grows, AssemblyAI or Deepgram credits buy months of runway before any bill.

A few habits that saved me from surprises: read whether the free amount is per month, per day or one-time; check if billing has to be enabled; and look for per-channel or per-session billing rules.

Disclosure: I'm a co-founder of nopaywall, a free catalog that tracks free tiers and generation limits like these with a last-checked date on each card. The audio and API sections are at https://nopaywall.io/limits if you want to compare more providers.

Free AI video quotas worth knowing (Oct 2026 caps)

unicodefawn — Sat, 10 Oct 2026 08:41:53 +0000

Short list of free or freemium video-generation allowances that still show up when you need a clip for a demo or placeholder. Caps change often — treat these as rounded figures and check the provider's page.

I'm a co-founder of nopaywall, a public catalog of free AI offers, API tiers, generation limits, and alternatives. Cards there show the checked terms and a last-checked date.

One-time or small free credits

Runway — about 125 one-time free credits on the free plan; watermark on free exports (checked Oct 2026).
Replicate — about $5 in free credits for new accounts (checked Oct 2026).
Qiroq — about 10 credits per month at 480p with watermark (checked Oct 2026).

Monthly / recurring free lanes

Google Vids — about 6 clips per month on the free lane (checked Oct 2026).
Synthesia — free plan around 1,200 credits per month (roughly ~10 minutes of avatar video; check current plan terms) (checked Oct 2026).

Useful framing

"Free" almost always means a hard cap, a watermark, non-commercial rights, or a short trial. Eligibility can vary by country. If a card on nopaywall looks wrong, the catalog is updated from source pages on a schedule — open the card for the last-checked date.

More video + image + music caps: nopaywall.io/limits.

Free LLM API tiers in 2026: the actual rate limits, and why they keep changing

unicodefawn — Sat, 10 Oct 2026 00:55:51 +0000

Disclosure: I'm a co-founder of nopaywall, the catalog this post draws from.

If you've ever built a side project on a "free" LLM API, you know the problem: the free tier is real, but the limits are buried in docs, they change without notice, and the blog post you found last year is wrong now. I'll go through what the free tiers look like at the moment and how we keep the data current.

The numbers (as listed on 2026-10-07, rounded)

Provider	Free limit	Notes
Groq	~30 RPM, ~1K requests/day, ~8K tokens/min	GPT-OSS 120B/20B, Qwen, Whisper
Google AI Studio (Gemini API)	~15 RPM, ~500 requests/day, ~250K tokens/min
Mistral Codestral	~30 RPM, ~2K requests/day	code-focused
Hugging Face Inference	~300 requests/hour
Puter.js	~30 requests per 10 s, 3 concurrent	runs from the browser

These are caps, not promises. Eligibility can vary by country, so check each provider's terms before you ship anything that depends on them.

Practical takeaways

Prototype vs. production. Around 500 to 1K requests a day is enough for a demo, a personal tool or a hackathon. A public app with real users will hit it on day one.
Watch tokens per minute, not just RPM. An 8K TPM limit allows only a few long-context calls per minute, even if you're well under the request cap.
Build a fallback chain. Since several providers serve the same open model families (Llama, Qwen, DeepSeek, Mistral, Gemma, GPT-OSS…), you can put an OpenAI-compatible router in front and fail over when you hit a 429.
Pin the date. When you write down a limit in your README, write the date you checked it next to it.

How the catalog stays current

nopaywall lists 100+ free API providers, with a model finder so you can search by model family instead of by vendor. Collection and checking are automated with source evidence: sources (GitHub repos and topic lists) are collected, then the official pricing or docs page is inspected. A card goes public only when it has a reachable destination and enough evidence for the limit. API providers and limits are re-checked within 72 hours. Cards are removed after a confirmed expiry or a failed verification. Unclear records stay off public pages.

Free, freemium, open-source and paid access are labeled separately. That way "free trial, card required" doesn't get mixed up with "free tier".