Higgs Audio V2.5 (TTS) API Documentation

Higgs Audio V2.5 is a 1B parameter autoregressive audio transformer distilled from the 3B V2 model. Featuring the DualFFN architecture and trained on 1M+ hours of AudioWeb data with GRPO alignment, it delivers human-like naturalness and low-latency speech synthesis.

This endpoint follows the OpenAI-compatible /v1/chat/completions schema, outputting high-fidelity 24kHz audio via a unified 25 FPS tokenizer.

---

Billing

$0.045 / audio minute

Billed based on the total duration of generated audio.

---

Parameters Detail

Request Body (JSON)

Parameter	Type	Required	Description
model	string	Yes	Must be "boson/higgs-audio-v2.5".
messages	array	Yes	Chat-formatted input. The model generates audio based on the last user message.
modalities	array	Yes	Must include ["text", "audio"] to enable TTS mode.
audio	object	No	Specifies audio details such as voice (e.g., alloy, echo) and format (wav, mp3).
temperature	float	No	Controls randomness in prosody and naturalness. Recommended: 1.0.
top_p	float	No	Nucleus sampling threshold. Recommended: 0.95 for optimal acoustic stability.
max_completion_tokens	integer	No	Limits the total tokens generated (indirectly limits audio duration).
extra_body	object	No	Model-specific parameters, such as {"top_k": 50} for decoding efficiency.
stop	array	No	Stop sequences, e.g., `["<

---

Response Fields (verbose_json)

Top-level Object

Field	Type	Description
id	string	Unique identifier for the request.
object	string	Always "chat.completion".
created	integer	Unix timestamp of the request.
model	string	The exact model version executed.
choices	array	List of generated outputs (typically contains one item).
usage	object	Statistics including prompt_tokens and completion_tokens.

choices[].message.audio

Field	Type	Description
data	string	Base64-encoded audio data. Default sampling rate is 24kHz.
id	string	Unique ID for the audio resource.
expires_at	integer	Expiration timestamp for the audio data (if cached).
transcript	string	The normalized text content corresponding to the generated audio.