dsh-tts

dsh-tts

Text-to-speech plugin for dsh: one tts tool that turns text into an audio file (wav / mp3 / flac / aac / opus / pcm) through any OpenAI-compatible TTS backend, so agents can voice their output without per-backend glue. The same tool lists voices and models, probes the routes, and calls the backend's dev phonemize endpoint.

Tool verbs

Key synth behaviors:

• The voice argument is required, and there is no default: the voice set is provider-specific. The available ids are listed by the voices verb.
• When the call omits model, the tool uses the route's single declared model. When a route declares no models, the tool resolves the model by a live /models probe. When the declared or probed model list contains several models, the call fails and lists the candidates.
• The response_format argument defaults to wav, and the tool sends the value verbatim.
• The backend owns container correctness, and the tool validates the returned bytes: a RIFF walk for wav, and magic checks for the other formats (pcm carries no container check). When the payload is not the claimed container, the call fails and the tool keeps the file on disk. When the container carries a known defect, for example the placeholder RIFF chunk sizes (0xFFFFFFFF or 0x7FFFFFFF) some backends emit, the successful document reports it in warnings.
• When the call sets out, the output path is that argument (a file path, or a directory marked by a trailing separator or by existing on disk). When the call omits out, the tool writes the file to <process.cwd()>/tts/<unixts>_<voice>.<ext>. The file extension comes from the response_format value.
• The speed and language arguments are sent verbatim, and only when the call sets them.
• The synth timeout is 180 s, and the probe and phonemize timeouts are 15 s and 30 s.

The plugin ships a dsh-tts skill. The skill is the full tool reference for agent contexts: it carries the parameter table, the env contract, and the error shapes.

Installation

dsh-tts is an out-of-tree plugin for a dsh profile. Install it with the dsh plugin command:

dsh plugin --profile <name> add dsh-tts

Configuration is layered: built-in defaults, the TTS_* env contract, a cordis patch file, a dsh-tts settings section (see below); the plugin creates no config or state files.

Configuration

Providers are configured in any of four layers, lowest to highest priority; the providers verb reports the originating layer of every field, so a shadowed value is never silent:

1. the built-in default (api: openai-compatible),
2. the TTS_* env contract (static per process boot),
3. the cordis config layer (a patch entry with id dsh-tts and a config object, or profile-level patches),
4. the dsh-tts: section in the user settings document ($DSH_HOME/settings.yaml), which the plugin re-reads on every call, so settings edits apply without a restart.

dsh-tts:
  defaultProvider: local
  providers:
    local:
      baseURL: http://<host>:<port>/v1
    hosted:
      baseURL: https://api.example.com/v1
      apiKeyEnv: TTS_HOSTED_KEY
      models: [kokoro-v2]
      displayName: Hosted Speech

The env contract mirrors the config fields 1:1: TTS_DEFAULT_PROVIDER, and per route <NAME> the variables TTS_PROVIDER_<NAME>_BASEURL, TTS_PROVIDER_<NAME>_API, TTS_PROVIDER_<NAME>_APIKEYENV, TTS_PROVIDER_<NAME>_HEADERS, TTS_PROVIDER_<NAME>_MODELS, TTS_PROVIDER_<NAME>_DISPLAYNAME. The <NAME> segment is the route name uppercased with - written as _. The headers value is a JSON object string, and the models value is a comma-separated list of model ids. The baseURL value is the OpenAI-compatible base URL including the /v1 prefix (the same shape as an LLM provider base URL); the plugin appends /audio/speech, /audio/voices, and /models to it verbatim, and it addresses /health and /dev/phonemize at the server root, which it derives by stripping a trailing /v1 from the base URL. The apiKeyEnv value is a reference to an environment variable name, never a secret in config; per request the value resolves from the process environment, then from $DSH_HOME/.credentials.yaml.

Examples

// Synthesize with the default route and its single declared model.
{ "verb": "synth", "input": "hello world", "voice": "zf_xiaoxiao" }
// Pick a route, format, speed, and output location explicitly.
{ "verb": "synth", "provider": "hosted", "input": "the quick brown fox",
  "voice": "zf_xiaoxiao", "response_format": "mp3", "speed": 1.1, "out": "/tmp/hi.mp3" }
// Find the voices by an id prefix filter.
{ "verb": "voices", "filter": "zf_" }
// Or by a case-insensitive substring over the id and name.
{ "verb": "voices", "query": "zf" }
// List what is configured, and where each field comes from.
{ "verb": "providers" }
// Probe the backend health (the live /models probe).
{ "verb": "status" }
// Ask the backend dev phonemize endpoint (a 404/405 response reports it as unsupported).
{ "verb": "phonemize", "text": "hello world", "language": "zh" }

Troubleshooting

A failing call returns a JSON document with an error message field, and the message text identifies the failure.

• When the error message is "TTS_NOT_CONFIGURED", no route survived configuration. Check the base URLs; the tool reports each route that configuration drops as a warning in the log.
• When the error message carries the text "credential missing: environment variable "X" is not set", the credential reference does not resolve. Set X in the environment or add it to $DSH_HOME/.credentials.yaml.
• When the error message carries the text "unknown provider "x"", the route name matches no configured route. Route names match case-insensitively and treat _ and - as the same separator; the error lists the configured names.
• When the error message carries the text "multiple providers configured (...) and no default is set", pass the provider argument explicitly, or set TTS_DEFAULT_PROVIDER, or set the settings defaultProvider.
• When status reports reachable: false, the error field carries the cause; the probe uses a 15 s budget against GET {baseURL}/models.
• When voices reports source: static, the backend answered 404 to GET {baseURL}/audio/voices, and the tool returned the static OpenAI table. The listed voice ids may not exist on that backend.
• When the synth document carries a container check error, the tool could not verify the returned bytes as the claimed container. Inspect the kept file at out; the backend may return an unsupported container for the requested response_format.

Build

pnpm install
pnpm run build     # tsc -> lib/
pnpm run lint      # eslint src/ test/
pnpm run selftest  # in-process: HTTP stub + config matrix + container checks