realtime-voice-agent
Operate MOSS-TTS-Realtime low-latency voice-agent streaming, Gradio, FastAPI, sessions, text deltas, and reference prompts.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- realtime-voice-agent
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Operate MOSS-TTS-Realtime low-latency voice-agent streaming, Gradio, FastAPI, sessions, text deltas, and reference prompts.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md9.9 KB
- 参考文档references/fastapi-service.md7.0 KB
- 参考文档references/streaming-workflows.md11.7 KB
- 参考文档references/troubleshooting.md8.7 KB
- 脚本scripts/realtime_session_payloads.py5.4 KB
03 SKILL.md 正文
realtime-voice-agent
Use this sub-skill when the task is about MOSS-TTS-Realtime voice-agent workflows: low-latency conversational TTS, reference-prompt voice cloning, Gradio streaming demos, FastAPI streaming sessions, client payloads, text-delta bridges from an LLM, or single-turn and multi-turn Python streaming with KV-cache control.
Route elsewhere
- MOSS-TTS-Local-Transformer-v1.5 local streaming, 48 kHz stereo codec-v2, continuation prompts, or v1.5 browser app behavior:
../local-v15-streaming/SKILL.md. - Generic Hugging Face model loading, non-Realtime family generation, model packaging, or broad HF API questions:
../hf-family-workflows/SKILL.md. - Fine-tuning, manifest preparation, dataset conversion, or training workflows:
../finetuning-data-prep/SKILL.md.
Minimum operating facts
- Realtime checkpoint:
OpenMOSS-Team/MOSS-TTS-Realtime. - Audio codec/tokenizer checkpoint:
OpenMOSS-Team/MOSS-Audio-Tokenizer. - Audio sample rate: 24 kHz mono for Realtime generation and service output.
- Recommended decoding defaults: temperature
0.8, top-p0.6, top-k30, repetition penalty1.1, repetition window50. - Streaming prefill normally waits for
12stable text tokens unless the stream is finalized earlier. - The bundled FastAPI service supports request/audio split sessions and currently assumes batch size
1.
Use the bundled references
- End-to-end local workflows, basic inference, Gradio, single-turn streaming, multi-turn KV-cache reuse, text deltas, and prompt audio:
references/streaming-workflows.md. - FastAPI server launch, endpoints, session flow, client behavior, response headers, and operational caveats:
references/fastapi-service.md. - Public classes, method contracts, constants, JSON schemas, and decoding defaults:
references/api-reference.md. - Failure symptoms and fixes for prompt audio, empty streams, finalization, KV cache, token boundaries, batch size, server config, downloads, attention kernels, and memory:
references/troubleshooting.md. - Safe payload planner with no server/model imports:
scripts/realtime_session_payloads.py.
Fast path
- For Python streaming, read
references/streaming-workflows.mdand choose single-turn, multi-turn, or text-delta bridge. - For HTTP service use, read
references/fastapi-service.md, start the server on the intended GPU, then generate payloads withscripts/realtime_session_payloads.py. - Before debugging audio quality or latency, check
references/api-reference.mdfor the exact session and decoder contracts. - If no audio arrives, a request hangs, or speaker identity changes across turns, go directly to
references/troubleshooting.md.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).