import mlx.core as mx
from mlx_audio.sts.models.lfm_audio import (
    LFM2AudioModel,
    LFM2AudioProcessor,
    ChatState,
    LFMModality,
)
from mlx_audio.sts.models.lfm_audio.model import AUDIO_EOS_TOKEN

# Load model and processor
model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")
processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")

# Create chat state
chat = ChatState(processor)
chat.new_turn("system")
chat.add_text("Perform TTS. Use a UK male voice.")
chat.end_turn()
chat.new_turn("user")
chat.add_text("Hello, welcome to MLX Audio!")
chat.end_turn()
chat.new_turn("assistant")

# Generate with interleaved text and audio
audio_codes = []
for token, modality in model.generate_sequential(
    **dict(chat),
    max_new_tokens=2048,
    temperature=0.8,

):
    mx.eval(token)
    if modality == LFMModality.AUDIO_OUT:
        if token[0].item() == AUDIO_EOS_TOKEN:
            break
        audio_codes.append(token)

# Decode audio
audio_codes = mx.stack(audio_codes, axis=0)[None, :].transpose(0, 2, 1)
waveform = processor.decode_audio(audio_codes)

# Save audio (24kHz sample rate)
import soundfile as sf
sf.write("output.wav", waveform[0].tolist(), model.sample_rate)

Speech-to-Text (ASR)

python

import mlx.core as mx
import numpy as np
import soundfile as sf
from mlx_audio.sts.models.lfm_audio import (
    LFM2AudioModel,
    LFM2AudioProcessor,
    ChatState,
    LFMModality,
)

# Load model and processor
model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")
processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")

# Load audio (must be 24kHz for audio input)
audio, sr = sf.read("input.wav")
audio = mx.array(audio.astype(np.float32))

# Create chat state with audio input
chat = ChatState(processor)
chat.new_turn("user")
chat.add_audio(audio, sample_rate=sr)
chat.add_text("Transcribe the audio.")
chat.end_turn()
chat.new_turn("assistant")

# Generate text response
text_out = []
for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=512):
    mx.eval(token)
    if modality == LFMModality.TEXT:
        text_out.append(token)
        print(processor.decode_text(token[None]), end="", flush=True)

Speech-to-Speech (STS)

python

import mlx.core as mx
import numpy as np
import soundfile as sf
from mlx_audio.sts.models.lfm_audio import (
    LFM2AudioModel,
    LFM2AudioProcessor,
    ChatState,
    LFMModality,
)

# Load model and processor
model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")
processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-5bit")

# Load input audio (24kHz)
audio, sr = sf.read("input.wav")
audio = mx.array(audio.astype(np.float32))

# Create chat state with audio input
chat = ChatState(processor)
chat.new_turn("system")
chat.add_text("Respond with interleaved text and audio.")
chat.end_turn()
chat.new_turn("user")
chat.add_audio(audio, sample_rate=sr)
chat.end_turn()
chat.new_turn("assistant")

# Generate response with both text and audio
text_out, audio_out = [], []
for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
    mx.eval(token)
    if modality == LFMModality.TEXT:
        text_out.append(token)
        print(processor.decode_text(token[None]), end="", flush=True)
    else:
        audio_out.append(token)

# Decode audio response
if audio_out:
    audio_codes = mx.stack(audio_out[:-1], axis=1)[None, :]  # (1, 8, T)
    waveform = processor.decode_with_detokenizer(audio_codes)
    sf.write("response.wav", waveform[0].tolist(), 24000)

Interleaved Text and Audio Generation

LFM2.5-Audio uses generate_interleaved for mixed text and audio output. The model can respond with text, audio, or both interleaved together.

Each audio token returned by generate_interleaved is a complete frame of shape (8,) containing all 8 codebook values:

python

from mlx_audio.sts.models.lfm_audio import LFMModality

text_out, audio_out = [], []
for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
    mx.eval(token)
    if modality == LFMModality.TEXT:
        text_out.append(token)
        # Stream text output
        print(processor.decode_text(token[None]), end="", flush=True)
    else:  # LFMModality.AUDIO_OUT
        audio_out.append(token)  # token shape: (8,)

# Stack audio frames: list of (8,) -> (8, T)
if audio_out:
    audio_codes = mx.stack(audio_out[:-1], axis=1)[None, :]  # (1, 8, T)
    waveform = processor.decode_with_detokenizer(audio_codes)

Audio Decoding Options

LFM2.5-Audio supports two methods for decoding audio codes to waveforms:

1. Detokenizer (Recommended for TTS)

The neural detokenizer reconstructs audio using ISTFT from predicted spectrograms:

python

# Decode using detokenizer
audio = processor.decode_with_detokenizer(codes[None])  # (1, T_audio)

2. Mimi Codec

The Mimi neural codec provides an alternative decoding path:

python

# Decode using Mimi codec
audio = processor.decode_audio(codes)  # (1, 1, T_audio)

Generation Configuration

python

from mlx_audio.sts.models.lfm_audio import GenerationConfig

config = GenerationConfig(
    max_new_tokens=2048,    # Maximum tokens to generate
    temperature=0.9,        # Text sampling temperature
    top_k=50,               # Text top-k sampling
    top_p=1.0,              # Text nucleus sampling
    audio_temperature=0.7,  # Audio sampling temperature
    audio_top_k=30,         # Audio top-k sampling
)

Streaming Generation

For real-time audio playback during generation:

python

from mlx_audio.sts.models.lfm_audio import LFMModality

FRAMES_PER_CHUNK = 10  # Decode every 10 audio frames

audio_buffer = []
for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
    mx.eval(token)
    if modality == LFMModality.AUDIO_OUT:
        audio_buffer.append(token)

        # Decode when we have enough frames
        if len(audio_buffer) >= FRAMES_PER_CHUNK:
            codes = mx.stack(audio_buffer, axis=1)[None, :]  # (1, 8, T)
            chunk = processor.decode_with_detokenizer(codes)
            # Play chunk with your audio library...
            audio_buffer = []

    elif modality == LFMModality.TEXT:
        # Stream text output
        print(processor.decode_text(token[None]), end="", flush=True)

Model Architecture

LFM2.5-Audio consists of:

Audio Encoder: Conformer-based encoder for processing input audio
LFM Backbone: 1.5B parameter Liquid Foundation Model for multimodal reasoning
Audio Decoder: Depthformer for generating audio codes
Detokenizer: ISTFT-based neural vocoder for waveform reconstruction

API Reference

LFM2AudioModel

python

class LFM2AudioModel:
    @classmethod
    def from_pretrained(cls, model_name: str) -> "LFM2AudioModel":
        """Load pretrained model from HuggingFace Hub."""

    def generate_interleaved(
        self,
        text_tokens: mx.array,
        audio_features: mx.array,
        modalities: mx.array,
        max_new_tokens: int = 512,
        temperature: float = 0.9,
        audio_temperature: float = 0.7,
        audio_top_k: int = 30,
    ) -> Generator[Tuple[mx.array, LFMModality], None, None]:
        """Generate interleaved text and audio tokens.

        Yields:
            (token, modality) tuples where:
            - For TEXT: token is scalar, modality is LFMModality.TEXT
            - For AUDIO_OUT: token is (8,) array, modality is LFMModality.AUDIO_OUT
        """

LFM2AudioProcessor

python

class LFM2AudioProcessor:
    @classmethod
    def from_pretrained(cls, model_name: str) -> "LFM2AudioProcessor":
        """Load pretrained processor from HuggingFace Hub."""

    def preprocess_audio(self, audio: mx.array, sample_rate: int) -> mx.array:
        """Convert audio to mel spectrogram features."""

    def tokenize_audio(self, audio: mx.array, sample_rate: int) -> mx.array:
        """Tokenize audio using Mimi codec."""

    def decode_audio(self, codes: mx.array) -> mx.array:
        """Decode audio codes using Mimi codec."""

    def decode_with_detokenizer(self, codes: mx.array) -> mx.array:
        """Decode audio codes using neural detokenizer."""

    def tokenize_text(self, text: str) -> mx.array:
        """Tokenize text."""

    def decode_text(self, tokens: mx.array) -> str:
        """Decode text tokens."""

ChatState

python

class ChatState:
    def __init__(self, processor: LFM2AudioProcessor):
        """Initialize chat state."""

    def new_turn(self, role: str):
        """Start a new turn (user/assistant/system)."""

    def end_turn(self):
        """End the current turn."""

    def add_text(self, text: str):
        """Add text to current turn."""

    def add_audio(self, audio: mx.array, sample_rate: int):
        """Add audio to current turn."""

License

This implementation follows the license terms of the original LFM2.5-Audio model. See LiquidAI/LFM2.5-Audio-1.5B for details.