Your voice.
In everyone's language.

Models recognize more than 60 languages ​​or 55 languages ​​in multilingual mode, switch languages ​​within speech and mark the language of words. The medical model covers 10 languages; pharmaceutical works with English drug names and dosages. Context includes names, subject area, terms, and preferred translations; the list of key terms contains up to 100 terms or 500 tokens and is updated during the session.

The text receives punctuation, formatting for numbers, dates, currencies, phone numbers and addresses. Confidence scores from 0 to 1 and timestamps are available for words and text. Medical batch recognition claims 97.20% term accuracy; improvement by keywords - up to 90% according to the provider.

Diarization distinguishes voices, separate processing distinguishes audio channels. Streaming and batch diarization use different versions. Determining speech activity and the semantic end of a phrase controls boundaries; the typical threshold for one model is 10 ms. Hiding of personal and payment data, entity extraction, topics, intents, sentiment and summaries with language and mode restrictions are available; Streaming entity extraction works in English.

Depending on the mode, file streaming and processing are provided. For one plan, 150 streaming and 50 file requests are specified simultaneously; The individual diarization limits are 50 for North America and 25 for Europe, Australia and India. The current room transmits 16 kHz mono PCM with the speaker language explicitly selected.

Terminology glossaries and translation memory help preserve the accepted vocabulary. The context sets the topic without translating the context itself; adaptive translation uses parallel examples. Formality settings, style rules and up to 10 custom instructions of 300 characters are available. The style profile combines vocabulary, style, and memory; separate instructions do not apply simultaneously.

Models offer speed or quality priority modes, domain models, and name translation options. Dictionary examples and back translations are available for a limited set of language pairs involving English.

Regional language options, transliteration and romanization, preservation of HTML/XML and protected elements, and control of sentence breaks and formatting are available. Experimental matching between source and translation is limited to language pairs and does not work with HTML. Processing of profanity allows for retention, flagging, or deletion.

DOCX, PDF, XLSX and PPTX documents are translated one at a time or in batches; one package supports up to 100 files and 10 target languages. Query options include up to 128 lines or up to 1,000 items and 50,000 characters with summation by target. For adaptive mode, the quota is 1 request per second per project. Translation works with individual text fragments and documents.

Model catalogs offer ready-made voices and filters by language, region, accent, pitch and personality. Controls of tempo, volume, intonation, emotion and pauses are available; select models support two voices in dialogue, delivery tags, phonetic notation and pronunciation dictionaries.

Creating a voice by description and cloning by sample are available in certain models and plans. For fast cloning, a sample from 10 to 60 seconds, a file up to 16 MB; professional training - from 30 minutes of audio and up to 3 hours. The other mode requires 10–30 seconds and a recording of consent.

Streaming output supports sending text in parts, canceling and continuing the context; timestamps of words, symbols or phonemes are available in separate modes. PCM, WAV, MP3, Opus, FLAC and telephone codecs are supported; frequencies from 8 to 48 kHz are determined by the format. Some models accept up to 40,000 characters per request; others - 4,096 characters or 2,000 input tokens.

For the models, the stated first response is 20 ms at P90 without network latency, about 75 ms of calculation or about 90 ms to the first byte; these are not Attuni latency measurements. Streaming audio uses mono PCM 24 kHz.