Published by TonePerfect, one of the APIs below. Vendor facts come from their public pages, checked on 10 September 2026; free-tier terms change often, so confirm on the vendor's site before you plan around them. Trial sizes we could not verify are not quoted.
At a glance
| Option | What is free | Card needed | After the free part | The catch |
|---|---|---|---|---|
| Azure AI Speech | A free (F0) Speech tier with a monthly allowance of audio hours | An Azure account | About $1 per audio hour, standard real-time | Full results through the Speech SDK; thresholds are yours to tune |
| TonePerfect | 50 credits per account through the real API, plus a live demo with no account | No | $10/month for 2,500 credits | One-time credits, not monthly; scripted speech only |
| SpeechSuper | A trial | See vendor | $0.004–0.008 per request, $20/month minimum | Trial size not published |
| Speechace | A free trial on its plans | See vendor | From $40/month (5,000 requests) | Fluency on Pro, unscripted on Premium |
| ELSA API | No free assessments listed | — | Price on request | English only; talk to sales |
| Open source (GOP, wav2vec 2.0) | The code and models | No | Your own servers and time | You build calibration, decisions and an API |
Azure AI Speech: the only recurring free tier
Microsoft bills pronunciation assessment as speech-to-text time, and its free (F0) Speech tier includes a monthly allowance of audio hours. For short scripted clips that allowance goes a long way, and the paid rate — about $1 per audio hour on the standard real-time tier — is the lowest per clip of any vendor here. Azure lists 33 locales for pronunciation assessment, unscripted and streaming modes, and prosody on en-US. What you give up: the complete feature set lives in the Speech SDK, results need parsing, and deciding when a score is low enough to correct a learner is up to you. Mandarin is scored per phoneme, not as initial, final and tone. TonePerfect vs Azure →
TonePerfect: 50 free credits and a demo without an account
Our own API, so judge accordingly. Every new account gets 50 free credits — up to 50 recordings of 15 seconds — that work through the real POST /v1/assess endpoint with your own API key, with no card and no subscription. Before signing up you can record yourself in the live demo on the homepage in any of the 10 languages and see the real JSON. Failed and unassessable requests never use credits. The limits: the credits are a one-time grant, not a monthly allowance; free requests share a daily capacity limit (a free_capacity error asks you to retry the next day); and the API scores scripted speech only, up to 30 seconds per request. After that, plans start at $10/month for 2,500 credits. Quickstart →
SpeechSuper and Speechace: trials
Both offer trials rather than a free tier. SpeechSuper covers eight languages with word, sentence and paragraph products and an offline mobile SDK; paid use is per request ($0.004–0.008) with a $20 monthly minimum. Speechace is English-first (plus French and Spanish) with exam rubrics; plans start at $40/month for 5,000 requests and its trial comes with those plans. Neither publishes the exact trial allowance on the pages we checked, so ask before you plan a pilot around it. vs SpeechSuper → · vs Speechace →
Open source: free if your time is
Teams with ML staff can build scoring themselves. The classic route is goodness of pronunciation (GOP) on a forced aligner — Kaldi ships a GOP recipe for the openly available speechocean762 English learner corpus — and the modern one is a self-supervised encoder such as wav2vec 2.0 or HuBERT fine-tuned for phoneme recognition, with GOP-style scores on top. The models cost nothing; the work is in everything around them: a pronunciation dictionary per language, native-speaker calibration so natives are not flagged, a decision layer, robustness to phone microphones, and a service to host it. How pronunciation scoring works describes each of those steps.
How to test for free without fooling yourself
- Record 20 clips once: a few native speakers, a few strong learners, a few beginners with errors you know about, on the phones your users have.
- Run the same clips through every free option. The 50 TonePerfect credits and a trial elsewhere are enough for that.
- Check the natives first. An engine that flags native speech will annoy learners whatever its other scores.
- Check a wrong sentence. Read a different sentence from the reference; a good engine should score it near zero, not 60.
- Model the paid cost at your volume with the pricing comparison.