hexgrad: Kokoro 82M

Kokoro 82M is hexgrad’s open-weight text-to-speech model with eighty-two million parameters. Its model card describes a compact, decoder-only architecture based on StyleTTS 2 and iSTFTNet. Written text is converted to pronunciation information and then speech in a selected voice. The documented v1.0 release provides fifty-four voices across eight languages; that profile is distinct from the earlier English-only v0.19 release.

Select a matching language and voice

The voice catalogue covers American and British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese. The two English accents are variants of one language. Voice names and language settings belong together: af_heart is an American English example in the official workflow, while the catalogue has separate groups for the other languages and accents.

Choose the speaker with the actual material it will read. A short course narration can use one consistent preset voice across its sections; a multilingual collection can select a voice from each required language group. The catalogue’s coverage differs between groups. French, for example, is represented by one listed voice rather than the larger selection available for English.

Prepare pronunciation and passages

Kokoro’s reference pipeline converts written forms into phonemes. Its examples allow an explicit phonetic reading for a term, so a difficult name can be prepared rather than left to an ambiguous spelling. Keep the surrounding sentence natural and preserve the wording the listener should hear.

A practical passage is: “The guide has three parts. First, choose the material. Next, compare the examples. Finally, save the version you want to use.” Write complete phrases and keep the pauses implied by punctuation. For a long piece, organise the input into coherent paragraphs instead of treating the whole document as one uninterrupted utterance.

The voice guide says many speakers perform best with passages of about 100–200 tokens within an approximately 500-token span. It notes weaknesses on very short utterances, particularly below 10–20 tokens, and a tendency to rush longer passages, especially beyond 400 tokens. These are token-based preparation observations, not word-count or recording-duration limits. Combining short fragments or splitting a long passage can provide a more suitable unit of speech.

Audio output and pace

The official workflow yields the processed text, phonemes and an audio result for each generated segment. It saves those segments as WAV at 24,000 Hz and exposes a speed setting in the reference pipeline. Keep the segment text and order associated with the audio so the assembled narration follows the intended script.

A useful editing check is to listen across the boundary between two passages, then inspect recurring names and the pacing of lists. The audio output and phoneme view serve different purposes: one is the recording the listener hears, the other helps inspect how the written text is being pronounced.

Language coverage and useful limits

The maker explicitly notes that non-English support can be thin because pronunciation processing or voice-training data is limited. Its preset groups therefore should not be treated as equal guarantees of delivery for every passage. Choose a suitable speaker and review the real names and phrasing required by the task.

The useful deliverable is a spoken passage or assembled narration using a documented preset voice, with text prepared around its pronunciation and passage-length needs. This profile follows the maker’s v1.0 Kokoro 82M materials; another checkpoint, quantised implementation or externally modified voice has its own technical scope.