VocalNet-M2: Efficient Low-Latency SLM

• VocalNet-M2 is a low-latency spoken language model that integrates multi-codebook tokenization and multi-token prediction to eliminate the flow-matching bottleneck.
• The architecture achieves nearly 50% latency reduction (∼349 ms) and improves performance with a WER of 6.07% compared to previous models.
• Its modular Thinker–Talker pipeline with stacked MTP layers enables efficient real-time processing, ideal for voice assistants, human-robot interfaces, and telepresence.

VocalNet-M2 is a low-latency spoken LLM (SLM) that advances the architecture of end-to-end speech generation systems by introducing two core techniques: integrated multi-codebook tokenization and multi-token prediction (MTP). Developed with the aim of minimizing response latency in real-time interactive applications while maintaining high speech and text quality, VocalNet-M2 eliminates the flow-matching model bottleneck present in previous SLMs by directly generating multi-track speech tokens. The architecture is designed to support fast, efficient, and high-fidelity speech synthesis, making it applicable to use cases such as voice assistants, human-robot interaction, and telepresence systems (Wang et al., 13 Nov 2025).

1. Multi-Codebook Tokenization

VocalNet-M2 employs the XY-Tokenizer (Gong et al., 2025) for discretizing speech into multiple codebook representations. The tokenizer comprises $J=8$ independent codebooks, each containing $K$ embedding vectors of dimension $D$. For a given acoustic frame $x_t$, a small encoder network $E$ produces a latent vector $h_t=E(x_t)\in\mathbb{R}^D$. Each codebook $j$ then quantizes this latent via nearest-neighbor search:

$$a^{\mathrm{cb}, j}_t = \arg\min_{1\leq k\leq K}\|h_t - e^{(j)}_k\|^2_2, \qquad j=1,\ldots,8$$

This results in an 8-tuple of codebook indices at each timestep.

While the XY-Tokenizer's explicit loss is omitted in the primary text, the standard vector-quantization variational autoencoder (VQ-VAE) objective is implied:

$$\mathcal{L}_{\mathrm{VQ}} = \sum_t \sum_{j=1}^8 \left( \|\mathrm{sg}[h_t] - e^{(j)}_{a^{\mathrm{cb},j}_t}\|^2_2 + \beta\|h_t - \mathrm{sg}[e^{(j)}_{a^{\mathrm{cb},j}_t}]\|^2_2 \right), \quad 0<\beta\leq1$$

where $\mathrm{sg}$ denotes stop-gradient. The approach allows rich, low-latency, multi-channel acoustic representation, but the model’s robustness depends heavily on the diversity and quality of training data. Empirically, multi-codebook tokenization requires larger, higher-quality datasets to match the robustness and word error rate (WER) of single-codebook approaches.

2. Multi-Token Prediction (MTP) Acceleration

To address the latency introduced by serial, step-by-step autoregression in conventional speech decoders, VocalNet-M2 incorporates $K$0 MTP layers stacked atop its Talker module. Each MTP layer predicts the next future token in parallel, permitting inference to stride forward by $K$1 tokens per decoding pass.

Let $K$2 be the upsampled semantic representation, and let $K$3 sum previous codebook embeddings. The base Talker transformer $K$4 produces the next token set:

$K$5

Successive MTP layers predict tokens at $K$6 for $K$7:

$K$8

Training minimizes the total cross-entropy over next-step and future predictions:

$K$9

Empirical ablation demonstrates that $D$0 yields optimal performance (WER improves from 8.56% at $D$1 to 6.07% at $D$2, with stable UTMOS rating).

3. Architectural Design

VocalNet-M2 follows a modular Thinker–Talker pipeline:

1. Audio Encoding: Raw audio $D$3 is processed by a frozen Whisper-large-v3 encoder and a learned downsample adapter, producing representations $D$4.
2. Semantic Decoding (Thinker): The Thinker, a Qwen3-8B autoregressive transformer, generates $D$5 text tokens $D$6 and hidden states $D$7.
3. Fusion & Upsampling: Text token embeddings and Thinker hidden states are fused via a learned linear layer, then upsampled (by a factor of 3) to produce $D$8, matching the granular frame count.
4. Audio Token Generation (Talker with MTP): The Talker, an autoregressive transformer with eight parallel output heads, consumes $D$9 and previous codebook embeddings, outputting eight codebook indices per timestep. Stacked MTP layers predict future tokens in the same pass.
5. Vocoder Synthesis: The contiguous 8-codebook indices are fed directly to a lightweight neural vocoder, synthesizing the waveform without flow-matching.

Block Diagram:

Raw Audio → Whisper Encoder → Downsample Adapter ↓ Thinker (Qwen3-8B) ↓ Fusion Layer → Upsample ↓ Talker (8-track AR Transformer + $x_t$0 MTP Layers) ↓ 8 Codebook Token Streams ↓ Lightweight Neural Vocoder ↓ Output Audio Waveform

4. Empirical Evaluation and Comparative Performance

VocalNet-M2 was evaluated via a rigorous three-phase training regimen: (i) Talker pretraining on $x_t$110,000 h of TTS data (Emilia corpus), (ii) Downsample Adapter and Thinker training with LoRA on speech-to-text, and (iii) end-to-end fine-tuning on $x_t$27,000 h of speech-to-speech dialogues from VoiceAssistant, UltraChat, and Tulu-3-derived samples.

Core metrics:

• Text Quality: scored by AlpacaEval, Llama Questions, TriviaQA, Web Questions (0–10 scale)
• Speech Quality: UTMOS (predicted MOS), word error rate (WER, via Whisper-large-v3)
• Latency: first-chunk generation time (for 0.8 s of audio, single L20 GPU)

Summary Results:

Model	AlpacaEval	LlamaQ	TriviaQA	WebQ	WER	UTMOS	Latency (ms)
SLAM-Omni	3.50	2.94	0.39	0.84	5.78	4.46	702 ± 30
VocalNet-8B	7.12	7.95	6.24	6.48	3.64	4.49	556 ± 8
GLM-4-Voice	5.86	7.74	4.95	5.56	11.90	4.23	1060 ± 2
MiniCPM-o	6.13	7.72	6.43	7.16	9.52	4.14	894 ± 82
kimi-audio	6.49	8.10	6.15	7.10	14.71	2.87	1745 ± 140
Qwen2.5-Omni	6.01	7.90	5.89	6.88	2.31	4.34	—
VocalNet-M2	7.29	8.33	6.13	6.65	6.07	4.31	348.9 ± 2.9

VocalNet-M2 achieves a near 50% reduction in first-chunk latency (∼725 ms to ∼349 ms) relative to prior SLMs, retaining strong text and speech metrics.

Ablation Studies:

• Single vs. Multi-Codebook Tokenization: Multi-codebook models, when matched with high-quality, filtered data, close the WER and UTMOS gap with single-codebook models and provide a ∼2× latency speedup by forgoing the flow-matching model.
• Effect of MTP Layers: Increasing $x_t$3 improves WER up to $x_t$4, after which gains plateau.

5. Insights, Advantages, and Limitations

By directly generating multi-codebook tokens, VocalNet-M2 obviates the need for separate flow-matching decoders, constituting a principal source of latency reduction. The MTP stack enables significant stride in autoregressive decoding, further reducing inference cost, particularly in resource-constrained or real-time interactive scenarios.

A primary limitation emerges from the data requirements for multi-codebook tokenization. Achieving robust tokenization that rivals single-codebook systems in WER and audio quality demands both extensive and high-quality training data. Consequently, extensions such as adaptive codebook sizing, semi-supervised learning, and improved semantic-acoustic embedding mechanisms become pertinent directions for future research.

6. Application Domains and Prospects

VocalNet-M2's design, with first-chunk latency reduced to ∼350 ms, is particularly suited to real-time dialogue, voice-assistant interactivity, remote conferencing, and responsive human-robot interfaces. This latency regime supports near-natural conversational turn-taking and improved user experience in interactive systems.

A plausible implication is that the removal of the flow-matching constraint and the multi-token generation mechanism together set a new trajectory for scalable, efficient SLMs, provided that future methods can efficiently address the data brittleness currently observed in multi-codebook training paradigms (Wang et al., 13 Nov 2025).