BosonAI Higgs-Audio-V3-STT

Higgs-Audio v3 is a state-of-the-art Speech-to-Text (STT) foundation model developed by Boson AI. Unlike traditional ASR models, Higgs-Audio v3 is a "reasoning-capable" transcriber. It utilizes a 10B parameter architecture (optimized for inference) to not only convert speech into text but also to capture the nuances of speaker intent, emotion, and background context.

Key Capabilities

• Massive Multilingual Support: Superior accuracy across 94 languages with automatic language identification.
• Context-Aware Transcription: Outperforms Whisper-v3-Large by 20%+ in word error rate (WER) for complex, multi-speaker environments.
• Semantic Understanding: Can output structured summaries or sentiment analysis directly from the audio stream.
• Diarization & Timestamps: Built-in ability to distinguish between different speakers and provide word-level timestamps.
• Dual-FFN Architecture: Inherits the high-efficiency design from the Higgs series, allowing for faster-than-real-time processing on standard GPUs.

---

Request Parameters

To use the Higgs-Audio v3 STT model, send a POST request to the /v1/audio/transcriptions endpoint with the following parameters.

Parameter	Type	Required	Description
file	binary	Yes	The audio file to transcribe (mp3, wav, flac, m4a). Max 25MB.
model	string	Yes	Use "bosonai-higgs-audio-v3-stt".
language	string	No	ISO 639-1 code (e.g., en, zh). If omitted, auto-detection is used.
response_format	string	No	json, text, srt, vtt, or verbose_json.
prompt	string	No	Optional text to guide the model's style or terminology.

---

Optional Parameters (Advanced)

Parameter	Type	Default	Description
timestamp_granularities[]	array	["segment"]	Options: word, segment. Enables precise timing for each part.
temperature	float	0.0	Controls sampling randomness. 0.0 is recommended for high accuracy.
diarization	boolean	false	When true, the model labels different speakers (e.g., Speaker 1, Speaker 2).
sentiment_analysis	boolean	false	If enabled, includes speaker emotion