You can deploy this model and start earning money today!

bosonai/
higgs-tts-2-3b-base

Name: higgs-tts-2-3b-base
Availability: InStock

Views

No views yet

Higgs TTS 2: Redefining Expressiveness in Audio Generation

Check our open-source repository https://github.com/boson-ai/higgs-audio for more details!

Rename note: Higgs Audio V2 and Higgs Audio V2 Generation have been renamed to Higgs TTS 2.

We are open-sourcing Higgs TTS 2, a powerful audio foundation model pretrained on over 10 million hours of audio data and a diverse set of text data. Despite having no post-training or fine-tuning, Higgs TTS 2 excels in expressive audio generation, thanks to its deep language and acoustic understanding.

On EmergentTTS-Eval, the model achieves win rates of 75.7% and 55.7% over "gpt-4o-mini-tts" on the "Emotions" and "Questions" categories, respectively. It also obtains state-of-the-art performance on traditional TTS benchmarks like Seed-TTS Eval and Emotional Speech Dataset (ESD). Moreover, the model demonstrates capabilities rarely seen in previous systems, including automatic prosody adaptation during narration, zero-shot generation of natural multi-speaker dialogues in multiple languages, melodic humming with the cloned voice, and simultaneous generation of speech and background music.

Here's the demo video that shows some of its emergent capabilities (remember to unmute):

Here's another demo video that show-cases the model's multilingual capability and how it enabled live translation (remember to unmute):

Technical Details

Higgs TTS 2 adopts the "generation variant" depicted in the architecture figure above. Its strong performance is driven by three key technical innovations:

We developed an automated annotation pipeline that leverages multiple ASR models, sound event classification models, and our in-house audio understanding model. Using this pipeline, we cleaned and annotated 10 million hours audio data, which we refer to as AudioVerse. The in-house understanding model is finetuned on top of Higgs Audio v1 Understanding, which adopts the "understanding variant" shown in the architecture figure.
We trained a unified audio tokenizer from scratch that captures both semantic and acoustic features.
We proposed the DualFFN architecture, which enhances the LLM’s ability to model acoustics tokens with minimal computational overhead.

Audio Tokenizer

We introduce a new discretized audio tokenizer that runs at just 25 frames per second while keeping—or even improving—audio quality compared to tokenizers with twice the bitrate. Our model is the first to train on 24 kHz data covering speech, music, and sound events in one unified system. It also uses a simple non-diffusion encoder/decoder for fast, batch inference. It achieves state-of-the-art performance in semantic and acoustic evaluations. Check https://huggingface.co/bosonai/higgs-audio-v2-tokenizer for more information about the tokenizer.

Model Architecture -- Dual FFN

Higgs TTS 2 is built on top of Llama-3.2-3B. To enhance the model’s ability to process audio tokens, we incorporate the "DualFFN" architecture as an audio adapter. DualFFN acts as an audio-specific expert, boosting the LLM's performance with minimal computational overhead. Our implementation preserves 91% of the original LLM’s training speed with the inclusion of DualFFN, which has 2.2B parameters. Thus, the total number of parameter for Higgs TTS 2 is 3.6B (LLM) + 2.2B (Audio Dual FFN), and it has the same training / inference FLOPs as Llama-3.2-3B. Ablation study shows that the model equipped with DualFFN consistently outperforms its counterpart in terms of word error rate (WER) and speaker similarity. See our architecture blog for more information.

Evaluation

Here's the performance of Higgs TTS 2 on four benchmarks, Seed-TTS Eval, Emotional Speech Dataset (ESD), EmergentTTS-Eval, and Multi-speaker Eval:

Seed-TTS Eval & ESD

We prompt Higgs TTS 2 with the reference text, reference audio, and target text for zero-shot TTS. We use the standard evaluation metrics from Seed-TTS Eval and ESD.

	SeedTTS-Eval		ESD
	WER ↓	SIM ↑	WER ↓	SIM (emo2vec) ↑
Cosyvoice2	2.28	65.49	2.71	80.48
Qwen2.5-omni†	2.33	64.10	-	-
ElevenLabs Multilingual V2	1.43	50.00	1.66	65.87
Higgs Audio v1	2.18	66.27	1.49	82.84
Higgs TTS 2 (base)	2.44	67.70	1.78	86.13

EmergentTTS-Eval ("Emotions" and "Questions")

Following the EmergentTTS-Eval Paper, we report the win-rate over "gpt-4o-mini-tts" with the "alloy" voice. Results of Higgs TTS 2 are obtained with the voice of "belinda". The judge model is Gemini 2.5 Pro.

Model	Emotions (%) ↑	Questions (%) ↑
Higgs TTS 2 (base)	75.71%	55.71%
gpt-4o-audio-preview†	61.64%	47.85%
Hume.AI	61.60%	43.21%
BASELINE: gpt-4o-mini-tts	50.00%	50.00%
Qwen 2.5 Omni†	41.60%	51.78%
minimax/speech-02-hd	40.86%	47.32%
ElevenLabs Multilingual v2	30.35%	39.46%
DeepGram Aura-2	29.28%	48.21%
Sesame csm-1B	15.96%	31.78%

^{_{'†' means using the strong-prompting method described in the paper.}}

Multi-speaker Eval

We also designed a multi-speaker evaluation benchmark to evaluate the capability of Higgs TTS 2 for multi-speaker dialog generation. The benchmark contains three subsets

two-speaker-conversation: 1000 synthetic dialogues involving two speakers. We fix two reference audio clips to evaluate the model's ability in double voice cloning for utterances ranging from 4 to 10 dialogues between two randomly chosen persona.
small talk (no ref): 250 synthetic dialogues curated in the same way as above, but are characterized by short utterances and a limited number of turns (4–6), we do not fix reference audios in this case and this set is designed to evaluate the model's ability to automatically assign appropriate voices to speakers.
small talk (ref): 250 synthetic dialogues similar to above, but contains even shorter utterances as this set is meant to include reference clips in it's context, similar to two-speaker-conversation.

We report the word-error-rate (WER) and the geometric mean between intra-speaker similarity and inter-speaker dis-similarity on these three subsets. Other than Higgs TTS 2, we also evaluated MoonCast and nari-labs/Dia-1.6B-0626, two of the most popular open-source models capable of multi-speaker dialog generation. Results are summarized in the following table. We are not able to run nari-labs/Dia-1.6B-0626 on our "two-speaker-conversation" subset due to its strict limitation on the length of the utterances and output audio.

	two-speaker-conversation		small talk		small talk (no ref)
	WER ↓	Mean Sim & Dis-sim ↑	WER ↓	Mean Sim & Dis-sim ↑	WER ↓	Mean Sim & Dis-sim ↑
MoonCast	38.77	46.02	8.33	63.68	24.65	53.94
nari-labs/Dia-1.6B-0626	-	-	17.62	63.15	19.46	61.14
Higgs TTS 2 (base)	18.88	51.95	11.89	67.92	14.65	55.28

Usage

Transformers 🤗

Higgs TTS 2 is supported natively in transformers: see the doc.

uv pip install "transformers>=5.3.0"

Single-speaker smart voice

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

conversation = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "Generate audio following instruction."}],
    },
    {
        "role": "scene",
        "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}],
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.",
            }
        ],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=False)
decoded = processor.batch_decode(outputs)
processor.save_audio(decoded, "output_single_speaker.wav")

Multi-speaker smart voice

Use [SPEAKER*] tags to generate a multi-speaker dialogue. Speaker characteristics are described in the scene role.

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

system_message = """You are an AI assistant designed to convert text into speech.
If the user's message includes a [SPEAKER*] tag, do not read out the tag and generate speech for the following text, using the specified voice.
If no speaker tag is present, select a suitable voice on your own."""

user_message = """[SPEAKER0] I can't believe you did that without even asking me first!
[SPEAKER1] Oh, come on! It wasn't a big deal, and I knew you would overreact like this.
[SPEAKER0] Overreact? You made a decision that affects both of us without even considering my opinion!
[SPEAKER1] Because I didn't have time to sit around waiting for you to make up your mind! Someone had to act."""

conversation = [
    {
        "role": "system",
        "content": [{"type": "text", "text": system_message}],
    },
    {
        "role": "scene",
        "content": [
            {"type": "text", "text": "Audio is recorded from a quiet room."},
            {"type": "text", "text": "SPEAKER0: feminine"},
            {"type": "text", "text": "SPEAKER1: masculine"},
        ],
    },
    {
        "role": "user",
        "content": [{"type": "text", "text": user_message}],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=2000, do_sample=False)
decoded = processor.batch_decode(outputs)
processor.save_audio(decoded, "output_multi_speaker.wav")

Zero-shot voice cloning

Clone a voice by providing a reference audio in the conversation history.

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

conversation = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "Generate audio following instruction."}],
    },
    {
        "role": "scene",
        "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}],
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "It was the night before my birthday. Hooray! It's almost here! It may not be a holiday, but it's the best day of the year.",
            }
        ],
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/eustlb/dummy-audio-samples-higgs/resolve/main/belinda.wav",
            }
        ],
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.",
            }
        ],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=False)
decoded = processor.batch_decode(outputs)
processor.save_audio(decoded, "output_voice_cloning.wav")

Multi-speaker voice cloning

Clone multiple voices by providing reference audio clips in the scene role.

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

user_message = """[SPEAKER0] I can't believe you did that without even asking me first!
[SPEAKER1] Oh, come on! It wasn't a big deal, and I knew you would overreact like this.
[SPEAKER0] Overreact? You made a decision that affects both of us without even considering my opinion!
[SPEAKER1] Because I didn't have time to sit around waiting for you to make up your mind! Someone had to act."""

conversation = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "Generate audio following instruction."}],
    },
    {
        "role": "scene",
        "content": [
            {"type": "text", "text": "Audio is recorded from a quiet room."},
            {"type": "text", "text": "SPEAKER0:"},
            {
                "type": "audio",
                "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/guess_age_gender.wav",
            },
            {"type": "text", "text": "SPEAKER1:"},
            {
                "type": "audio",
                "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac",
            },
        ],
    },
    {
        "role": "user",
        "content": [{"type": "text", "text": user_message}],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=False)
decoded = processor.batch_decode(outputs)
processor.save_audio(decoded, "output_multi_speaker_cloning.wav")

Batched inference

Process multiple conversations in a single forward pass.

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto")

conversation1 = [
    {"role": "system", "content": [{"type": "text", "text": "Generate audio following instruction."}]},
    {"role": "scene", "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}]},
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "It was the night before my birthday. Hooray! It's almost here! It may not be a holiday, but it's the best day of the year.",
            }
        ],
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/eustlb/dummy-audio-samples-higgs/resolve/main/belinda.wav",
            }
        ],
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.",
            }
        ],
    },
]

conversation2 = [
    {"role": "system", "content": [{"type": "text", "text": "Generate audio following instruction."}]},
    {"role": "scene", "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}]},
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": " It's super important to assess fairly the fact that our former model is over. And this is not a question of adjustment. This is not the same world, 2024, 2025. And on top of that, we are making the same mistakes, on top of the key elements I mentioned. We are over-regulating and under-investing. So just if, in the two to three years to come, if we follow our classical agenda, we will be out of the market. I have no doubts.",
            }
        ],
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/eustlb/dummy-audio-samples-higgs/resolve/main/macron.wav",
            }
        ],
    },
    {
        "role": "user",
        "content": [{"type": "text", "text": "Hey, here is a clone from the given voice."}],
    },
]

inputs = processor.apply_chat_template(
    [conversation1, conversation2],
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=False)
decoded = processor.batch_decode(outputs)
processor.save_audio(decoded, ["output_batched_1.wav", "output_batched_2.wav"])

Training

By default, the model does not load the text language modeling head to save memory (~1.5GiB reduction), as it's not required for generation. When training, set use_text_head=True to compute loss on text tokens.

python

from transformers import AutoProcessor, HiggsAudioV2ForConditionalGeneration

model_id = "bosonai/higgs-audio-v2-generation-3B-base"
processor = AutoProcessor.from_pretrained(model_id, device_map="auto")
model = HiggsAudioV2ForConditionalGeneration.from_pretrained(model_id, device_map="auto", use_text_head=True)

conversation1 = [
    {"role": "system", "content": [{"type": "text", "text": "Generate audio following instruction."}]},
    {"role": "scene", "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}]},
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "It was the night before my birthday. Hooray! It's almost here! It may not be a holiday, but it's the best day of the year.",
            }
        ],
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/eustlb/dummy-audio-samples-higgs/resolve/main/belinda.wav",
            }
        ],
    },
]

conversation2 = [
    {"role": "system", "content": [{"type": "text", "text": "Generate audio following instruction."}]},
    {"role": "scene", "content": [{"type": "text", "text": "Audio is recorded from a quiet room."}]},
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": " I would imagine so. A wand with a dragon heartstring core is capable of dazzling magic, and the bond between you and your wand should only grow stronger. Do not be surprised at your new wand's ability to perceive your intentions, particularly in a moment of need",
            }
        ],
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "audio",
                "url": "https://huggingface.co/datasets/eustlb/dummy-audio-samples-higgs/resolve/main/broom_salesman.wav",
            }
        ],
    },
]

inputs = processor.apply_chat_template(
    [conversation1, conversation2],
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    sampling_rate=24000,
    return_tensors="pt",
    output_labels=True,
).to(model.device)

outputs = model(**inputs)
outputs.loss.backward()

Original codebase

You need to first install the higgs-audio:

bash

git clone https://github.com/boson-ai/higgs-audio.git

cd higgs-audio
python3 -m venv higgs_audio_env
source higgs_audio_env/bin/activate
pip install -r requirements.txt
pip install -e .

Afterwards, try to run the following python code snippet to convert text to speech.

python

from boson_multimodal.serve.serve_engine import HiggsAudioServeEngine, HiggsAudioResponse
from boson_multimodal.data_types import ChatMLSample, Message, AudioContent

import torch
import torchaudio
import time
import click

MODEL_PATH = "bosonai/higgs-audio-v2-generation-3B-base"
AUDIO_TOKENIZER_PATH = "bosonai/higgs-audio-v2-tokenizer"

system_prompt = (
    "Generate audio following instruction.\n\n<|scene_desc_start|>\nAudio is recorded from a quiet room.\n<|scene_desc_end|>"
)

messages = [
    Message(
        role="system",
        content=system_prompt,
    ),
    Message(
        role="user",
        content="The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.",
    ),
]
device = "cuda" if torch.cuda.is_available() else "cpu"

serve_engine = HiggsAudioServeEngine(MODEL_PATH, AUDIO_TOKENIZER_PATH, device=device)

output: HiggsAudioResponse = serve_engine.generate(
    chat_ml_sample=ChatMLSample(messages=messages),
    max_new_tokens=1024,
    temperature=0.3,
    top_p=0.95,
    top_k=50,
    stop_strings=["<|end_of_text|>", "<|eot_id|>"],
)
torchaudio.save(f"output.wav", torch.from_numpy(output.audio)[None, :], output.sampling_rate)

You can also check https://github.com/boson-ai/higgs-audio/tree/main/examples for more example scripts.

License

See LICENSE

bosonai/higgs-tts-2-3b-base

Higgs TTS 2: Redefining Expressiveness in Audio Generation

Technical Details

Audio Tokenizer

Model Architecture -- Dual FFN

Evaluation

Seed-TTS Eval & ESD

EmergentTTS-Eval ("Emotions" and "Questions")

Multi-speaker Eval

Usage

Transformers 🤗

Original codebase

License

bosonai/
higgs-tts-2-3b-base