NeMo-Speech.cpp/docs/README.md at main · NVIDIA/NeMo-Speech.cpp

GitHub

Documentation

User and developer documentation for NeMo-Speech.cpp. The shortest path to a transcript lives in the

root README

; everything below is the detailed manual.

Start with:

Installation

Command-line workflows

Configuration examples

Model conversion

Speech-to-speech VoiceChat

HTTP/realtime and optional gRPC server

HTTP API reference

Client integration

Native SDK integration

Troubleshooting

Build from source

ASR

Models

- ready-to-run GGUF downloads, custom checkpoint conversion, and quantization options.

Configuration

- full asr.* key reference plus the decoding / word-boosting / VAD / endpointing / diarization / postprocessing knobs.

Feature matrix and customization

- which models support which features, and request-time vs startup customization.

TTS

Models

- ready-to-run MagpieTTS and NanoCodec GGUF downloads, tokenizer setup, and custom conversion.

Configuration

- tts.* key reference and serving.

NMT

Models and conversion

- obtain and convert Riva-Translate models for the llama.cpp runtime.

Configuration

- nmt.* key reference and serving.

S2S

Overview

- build and serve the streaming VoiceChat pipeline.

Models and conversion

- convert complete model repositories.

Docker deployment

- build and run the VoiceChat server image.

Configuration

- s2s.* settings and runtime tuning.

Client integration

- streaming, tool responses, and concurrency.

Developer guide

Overview

- implementation and performance internals.

Diagnostics

- check_backend_coverage.

ASR batching

- neural microbatching and streaming-state arenas.

ggml patches

- the project-specific ggml changes.

cuBLAS shim

- the in-tree drop-in cuBLAS replacement.