Canopy Labs: Orpheus 3B

Orpheus 3B 0.1 Fine-Tuned is Canopy Labs’ English text-to-speech model for everyday synthesis. It belongs to the Orpheus family built on a Llama 3B backbone and focuses on natural intonation, rhythm and expressive delivery. The input is the text to be spoken, with a selected speaker; the output is speech audio. The manufacturer also lists zero-shot voice cloning for this exact fine-tuned model. Its published fine-tuned synthesis example uses preset speakers, while reference-conditioned text–speech prompting is documented separately for the pretrained edition. This fine-tuned edition is distinct from the pretrained model and the separately released multilingual research models.

Choose a speaker and prepare the line

The manufacturer lists eight English preset voices: tara, leah, jess, leo, dan, mia, zac and zoe. Its prompt convention places the speaker name before the spoken text, separated by a colon. The reference speech-generation workflow can add that formatting when the voice is supplied separately. Select a speaker for the intended piece and keep that choice consistent across connected passages.

For a short explainer, write the final narration rather than an instruction asking the model to invent the explanation. A useful spoken line is: “Choose the folder you want to share, check who can open it, and save the changes.” For dialogue, prepare each speaker’s line explicitly and retain the punctuation that helps the delivery. An abbreviation, date or product name should be written in the form you want the listener to hear, particularly when its written shorthand would be ambiguous.

Expressive and nonverbal cues

Orpheus documents the tags <laugh>, <chuckle>, <sigh>, <cough>, <sniffle>, <groan>, <yawn> and <gasp>. They give the text an explicit nonverbal event at the relevant point. A short conversational passage can place a chuckle before a light remark or a sigh before a reflective sentence. Use the cue where it belongs in the performance and keep the actual words of the line clear.

These are speech-generation cues, not a general instruction language for editing an existing recording. For a voice-over, begin with readable copy and add only the expressive events the scene needs. Listen to the complete line to judge whether the event, emphasis and pacing support the meaning. A narration task may call for a steady voice, while a character line may benefit from a specific nonverbal cue.

Streaming and the audio result

The manufacturer supplies both a non-streaming notebook and a streaming speech example. Streaming returns successive audio chunks as the passage is generated. In the published example, those chunks are written into a mono WAV file at 24,000 Hz with 16-bit samples. Those file settings describe the reference workflow; the useful result is a playable recording whose words, speaker and delivery can be checked against the script.

For a longer narration, prepare coherent passages with clear sentence boundaries and keep a copy of the final script beside the audio. Review the start and end of each passage, the pronunciation of important names and the transition between adjacent recordings. The manufacturer’s sampling guidance specifies a repetition penalty of at least 1.1 for stable generation in its reference implementation and notes that sampling choices can affect speaking speed. The sample’s configured model length is an inference setting, rather than an established maximum narration duration.

Edition boundaries

This profile concerns the fine-tuned English model with its documented preset speakers. The pretrained model has its own text-and-speech conditioning workflow; multilingual releases have language-specific speakers and tags. Selecting this English edition does not combine those separate profiles. A practical deliverable is a spoken script, an intentional preset voice and audio that has been reviewed for the wording and performance required by the piece.