Questions and answers
What people ask before they plug OpenPronounce into a product. Each answer gives the figure when there is one, and says so when there is none yet.
Jump to: audio, hardware, scores, languages, data and licence.
Audio
Does it work with 8 kHz audio, or only 16 kHz?
Both, and any other rate. Every recording is converted to 16 kHz mono before analysis, so 8 kHz phone audio goes in as it is.
Phone audio costs a little. On a simulated land line (8 kHz, narrow band, G.711), the scores still follow the raters, but a given recording moves by 3 to 4 points on average compared with the same recording on a good microphone. Compare candidates recorded the same way, and record at 16 kHz or more whenever you control the device. The figures are in the list of limits.
From Python, assess expects a 16 kHz mono waveform: read files with load_audio, which does the conversion.
Which file formats does it read?
wav, mp3, flac and ogg directly, and anything else ffmpeg can decode: the webm/opus files that browsers record, m4a from phones. Stereo is mixed down to mono.
Is there a maximum length?
No. Free speech is cut at its pauses into parts of at most 12 seconds, each scored on its own, so a long recording simply takes longer: three minutes of audio take 52 seconds on a 16-thread CPU.
There are minimums. Under half a second, the recording is rejected as too short. In free speech, under 3 seconds of speech or four recognised words, the result comes back marked unreliable.
For read sentences, the length that matters is the sentence. The scorer learned on sentences of up to 10 words, and native speakers score lower from 20 words up. Split a paragraph into sentences and score each one.
What if two people speak on the same recording?
Both voices are scored together: OpenPronounce does not separate speakers. On a call, send only the channel of the person you are assessing, or cut out their turns before sending them.
What happens with a noisy or clipped recording?
It is scored anyway. audio_quality.issues names the problem (too quiet, clipped, noisy, too short) and reliable turns false. Your application decides whether to show the score or ask for another take.
Hardware
Do I need a GPU?
No. Everything runs on an ordinary CPU server. On 16 threads, a two-second sentence takes 0.45 seconds in quick mode, an 18-second reading 6.2 seconds with the full analysis, and three minutes of free speech 52 seconds.
What does a GPU change?
It is used automatically when PyTorch sees a CUDA card. OPENPRONOUNCE_DEVICE forces a choice: cpu, cuda, cuda:1, or mps on a Mac. The speech models run much faster on a GPU, but no end-to-end timings on one have been published yet, so measure on your own card before sizing a cluster around it.
How much memory and disk does it need?
About 4 GB of memory. The two speech models weigh 1.2 GB each and are downloaded on the first run; the Docker image already contains them.
How many recordings can one server handle?
A 16-thread CPU server analyses audio about three times faster than real time, close to 70 hours of audio a day. One server process handles one analysis at a time: for more, run several containers behind a load balancer.
Does it run on macOS and Windows?
On macOS, with pip and Homebrew's ffmpeg. On Windows, use Docker or WSL. The installation guide has the commands.
Can it run without network access?
Yes, once the speech models have been downloaded. Set HF_HUB_OFFLINE=1, and either use the offline Piper voices or quick mode, which needs no synthetic reading. See Configuration.
Scores
What does a score of 70 mean?
The scores reproduce the 0 to 10 ratings that five expert raters gave, multiplied by ten. These raters were strict: a clean adult reading gets about 80, not 100. Show a band rather than a decimal, because 76.6 and 78.1 describe the same performance.
How close is it to human raters?
On 2,500 test recordings whose speakers were kept out of training, the overall score correlates at 0.85 with the raters for adults reading a sentence, and 0.82 for adults speaking without a text. Averaged over ten recordings per speaker, the free-speech score ranks speakers at 0.88.
Your speakers will differ. openpronounce-evaluate gives the same figure on recordings your own assessors have scored. See Measure it on your data.
Does it penalise accents?
It follows its raters, who marked what hurt intelligibility and let mild accent traits pass. But it learned from one corpus of speakers, and accents far from theirs are less well covered. That is the first thing to check with your own recordings.
Will a native speaker get 100?
No. Natives reading score around 72 to 80. Natives talking freely score lower, about 66, because pauses and restarts count as faults, and very fast speech is not rewarded either.
What is quick mode, and what does it skip?
--fast on the command line runs the sound recogniser alone, without word transcription, synthetic reference reading or pitch curves. It answers in under half a second for a short sentence. The full analysis adds the transcript, the comparison with a native reading and the prosody curves.
Why does free speech report "unclear" words instead of mistakes?
Without the sentence, OpenPronounce cannot know which word the speaker meant. It can tell that a word came out unclear, not what it should have been. Pass the sentence whenever you have it.
Does it work with children?
With a sentence, yes, less well than with adults: 0.67 against 0.85. Without a sentence, no: 0.42. Have children read.
Does it judge grammar or content?
No. Pronunciation only: sounds, fluency and prosody. A well-pronounced wrong answer gets a good score.
Languages
Which languages does it support?
English, French, Spanish, German, Italian, Portuguese and Dutch. English is the only one measured against human raters. The six others are checked on native speakers, who score in them as they do in English, but no rated corpus exists yet to give a correlation.
Languages other than English also need the espeak-ng program. See Languages.
Can I add a language?
Possibly, if espeak-ng knows it. A measured version needs rated recordings in that language. If you have some, get in touch.
Data and licence
Do recordings leave our servers?
No. They are analysed on your machines and sent nowhere, so there is no vendor to add to your privacy policy and no data processing agreement to sign for them. There is no account, no licence server and no telemetry. One default sends data out: the first time a sentence is analysed, its text, never the audio, goes to Google's speech synthesis to build a native reading. OPENPRONOUNCE_TTS=piper or quick mode removes it.
Can we retrain the models on our own ratings?
Yes. The scoring models are JSON files, the training scripts are in the repository, and OPENPRONOUNCE_MODELS_DIR loads your versions in place of the shipped ones. openpronounce --models shows which are in use.
Is it free?
Yes, under the AGPL-3.0: code, models and scripts. A product whose own source code stays closed needs a yearly commercial licence. Testing and evaluating it never does. See Pricing.
A question not answered here? Ask it on GitHub, where the answer helps the next person, or use the contact page.