reference_language_code=zh to transcribe with a Chinese language hint and generate a Chinese preview. Omit it to retain automatic language detection. Speaker consent is required; see voice consent.
To create your own lines or cross-language speech, save the voice and use Text to Speech.
Set remove_background_noise=true to denoise a noisy sample before analysis (removeBackgroundNoise: true in TypeScript, --remove-background-noise in the CLI). The default is false. If denoising is unavailable, the original sample is used.
The Chinese reference option is referenceLanguageCode: "zh" in TypeScript, reference_language_code="zh" in Python, and --reference-language zh in the CLI. Chinese ASR failures return a retryable error; they never fall back to automatic language detection.
Sample requirements
- Upload exactly one sample file.
- Use a single speaker with no background music or noise.
- Use MP3, WAV, M4A, FLAC, AAC, OGG, Opus, WebM audio, WMA, or AIFF. The file must contain a decodable audio track; a supported extension does not guarantee every codec is supported.
- Minimum duration: 3 seconds. There is no maximum.
- Maximum file size: 5 MB.
filename; provide content_type for unnamed bytes that are not WAV. In TypeScript, supply a matching filename or contentType.
Step 1: Create a preview
BreezeBlue analyzes the first 60 seconds of the upload: it transcribes that stretch, detects its language, and keeps at most 30 seconds of it, ending on the last complete sentence that fits. Without usable punctuation the cut falls back to a natural pause, and failing that to the last whole word — it never lands mid-word. The excerpt starts just before the first word, keeping up to a second of leading silence, and is normalized to -18 LUFS so a quiet or hot recording does not bias the clone. The stored reference is that trimmed, normalized excerpt together with its transcript, not the original file. A long recording therefore clones as well as a tightly edited one: upload the whole take and let BreezeBlue pick the cut. Put the audio you want cloned in the first minute.generated_voice_id, not a saved voice_id. The preview script is generated by BreezeBlue in the recognized audio language, without performance instructions. SDKs also accept files with exactly one item. If the reference speech cannot be identified, upload a clearer sample in a supported language.
Step 2: Stream the preview
Step 3: Save the preview as a voice
language_code is required so later cross-language synthesis can distinguish the voice’s reference language from the requested speech language.
POST /v1/text-to-speech/{voice_id} calls use the stored reference sample and its transcript, not the preview audio.
Editing or deleting a saved voice
- Update tags and the description with
PATCH /v1/voices/{voice_id}. - Tune defaults with
PATCH /v1/voices/{voice_id}/settings. - Remove the voice with
DELETE /v1/voices/{voice_id}.
Continue building
Create clone preview
Upload one audio sample with an optional name and description.
Download completed voice preview
Listen to a
generated_voice_id preview before deciding to save it.Save voice preview
Persist the preview as a reusable voice; saving consumes a voice slot.
CLI voice clone
Clone from an audio sample in the terminal with
breeze voice clone.Voices
Learn how cloned voices, designed voices, and public voices fit together.
Text to speech
Use the saved cloned voice for dialogue lines, previews, or production audio.
Use the cloned voice
Save the preview, then use the savedvoice_id with Text to Speech to choose your script, instructions, and target language. The saved voice retains its original reference audio and transcript.
Names and reference analysis
An omitted, null, empty, or whitespace-only Clone name uses the audio filename stem, orCloned Voice when unavailable. Saving without a name keeps the preview name. Automatic names are limited to 80 Unicode characters; explicit names longer than 80 characters are rejected. Other preview types still require a name when saving.
A nonempty transcript in a supported main ASR language may proceed despite low confidence or mixed-language evidence. Unsupported languages and empty transcripts remain blocked. If denoising removes usable speech, analysis tries the original audio once within the same request budget; the selected audio and transcript stay together. Service errors remain retryable and are not audio-quality verdicts. File types, size, duration, model selection, and pricing are unchanged.
