MinMoOptions
Options & Configuration · AiDotNet.TextToSpeech.MultiModal
Options for MinMo.
Inheritance: ModelOptions → TtsModelOptions → MinMoOptions
For Beginners
MinMo is a voice assistant built on a text chatbot: an ear turns speech into something the chatbot reads, the chatbot writes its answer, and a voice reads the answer aloud as it is written.
How It Works
The defaults are the paper’s MinMo (Chen et al. 2025, Table 1): SenseVoice-Large’s encoder, an input projector of
one 2×-downsampling convolution and two Transformer layers, Qwen2.5-7B-Instruct with LoRA, a linear output projector
into CosyVoice 2’s voice token LM and its token2wav (VoiceDecoder), and a one-layer full duplex predictor.
Where the paper is silent the same team’s released code decides: FunASR’s LLM-ASR (Transformer adaptor, its
ChatML prompt, Qwen’s generation settings) and CosyVoice 2.
Constructors
| Constructor | Summary |
|---|---|
MinMoOptions() |
Creates the paper’s MinMo. |
Properties
| Property | Summary |
|---|---|
CosyVoice2Options VoiceDecoder { get; set; } |
CosyVoice 2 (0.5B), whose language model is the voice token LM and whose flow and vocoder are the token2wav synthesizer. |
MinMoTrainingStage Stage { get; set; } |
The stage a training step follows. |
MinMoVoiceEncoder Encoder { get; set; } |
The voice encoder. |
double LoraAlpha { get; set; } |
|
double LoraDropout { get; set; } |
|
double MaxGradientNorm { get; set; } |
Gradient clipping (not stated: 1.0). |
double RepetitionPenalty { get; set; } |
|
double RmsEpsilon { get; set; } |
|
double RopeTheta { get; set; } |
|
double Temperature { get; set; } |
Text generation: Qwen2.5-Instruct’s sampling (temperature 0.7, top-p 0.8, top-k 20, repetition penalty 1.05); a temperature of 0 decodes greedily. |
double TopP { get; set; } |
|
int DuplexFeedForward { get; set; } |
|
int DuplexHeads { get; set; } |
|
int DuplexWidth { get; set; } |
The full duplex predictor (18M): the LLM’s hidden states projected to 1024, one causal Transformer layer (8 heads, feed-forward 4096) and a linear softmax over MinMoDuplexAction. |
int Intermediate { get; set; } |
|
int KeyValueHeads { get; set; } |
|
int LargeEncoderHeads { get; set; } |
|
int LargeEncoderLayers { get; set; } |
|
int LargeEncoderMels { get; set; } |
SenseVoice-Large’s encoder: log-mel bands (128), width (1280), heads (20) and layers (32). |
int LargeEncoderWidth { get; set; } |
|
int Layers { get; set; } |
Qwen2.5-7B-Instruct: layers (28), key-value heads (4), feed-forward (18944), RoPE base (1e6) and RMSNorm epsilon (1e-6). |
int LoraRank { get; set; } |
The LoRA on the LLM’s query and value projections for instruction tuning (not stated: rank 8, alpha 16, dropout 0.05). |
int MaxNewTokens { get; set; } |
|
int ProjectorDownsample { get; set; } |
The input projector (FunASR’s Transformer adaptor): frames stacked by 2 (the 2× downsampling convolution), a 2048-wide ReLU bottleneck into the LLM’s width, then 2 Transformer layers of 8 heads whose feed-forward is a quarter of the width… |
int ProjectorFeedForward { get; set; } |
|
int ProjectorHeads { get; set; } |
|
int ProjectorLayers { get; set; } |
|
int SmallEncoderBlocks { get; set; } |
|
int SmallEncoderFeedForward { get; set; } |
|
int SmallEncoderHeads { get; set; } |
|
int SmallEncoderMels { get; set; } |
SenseVoice-Small’s SAN-M encoder: filterbank bands (80), width (512), heads (4), feed-forward (2048), blocks (50) and blocks after the norm (20). |
int SmallEncoderTpBlocks { get; set; } |
|
int SmallEncoderWidth { get; set; } |
|
int TopK { get; set; } |
|
string SynthesisInstruction { get; set; } |
|
string SystemPrompt { get; set; } |
The system prompt (FunASR’s LLM-ASR default). |
string TranscriptionInstruction { get; set; } |
The instruction Synthesize and Transcribe send (§3.3: “Speech Transcription” for recognition; Table 3’s form for synthesis). |