feat(llm): local Whisper backend via candle for offline STT

CHANGELOG.md

@@ -31,6 +31,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Telegram voice and audio message handling with automatic file download (#524)
 - STT bootstrap wiring: `WhisperProvider` created from `[llm.stt]` config behind `stt` feature (#529)
 - Slack audio file upload handling with host validation and size limits (#525)
+- Local Whisper backend via candle for offline STT with symphonia audio decode and rubato resampling (#523)
 - Shell-based installation script (`install/install.sh`) with SHA256 verification, platform detection, and `--version` flag
 - Shellcheck lint job in CI pipeline
 - Per-job permission scoping in release workflow (least privilege)

Cargo.toml

@@ -43,11 +43,13 @@ reqwest = { version = "0.13", default-features = false }
 rmcp = "0.15"
 scrape-core = "0.2.2"
 subtle = "2.6"
+rubato = "0.16"
 schemars = "1.2"
 similar = "2.7"
 serde = "1.0"
 serde_json = "1.0"
 serial_test = "3.3"
+symphonia = { version = "0.5.5", default-features = false, features = ["mp3", "ogg", "wav", "flac", "pcm"] }
 sqlx = { version = "0.8", default-features = false, features = ["macros"] }
 teloxide = { version = "0.17", default-features = false, features = ["rustls", "ctrlc_handler", "macros"] }
 tempfile = "3"

README.md

@@ -239,7 +239,7 @@ Skills **evolve**: failure detection triggers self-reflection, and the agent gen
 |----------|-------------|
 | **MCP** | Connect external tool servers (stdio + HTTP) with SSRF protection |
 | **A2A** | Agent-to-agent communication via JSON-RPC 2.0 with SSE streaming |
-| **Audio input** | Speech-to-text via OpenAI Whisper (25 MB limit); Telegram and Slack audio files transcribed automatically |
+| **Audio input** | Speech-to-text via OpenAI Whisper API or local Candle Whisper (offline, feature-gated); Telegram and Slack audio files transcribed automatically |
 | **Channels** | CLI, Telegram (text + voice), Discord, Slack, TUI — all with streaming support |
 | **Gateway** | HTTP webhook ingestion with bearer auth and rate limiting |
 | **Native tool_use** | Structured tool calling via Claude/OpenAI APIs; text fallback for local models |

crates/zeph-llm/Cargo.toml

@@ -10,7 +10,7 @@ repository.workspace = true
 default = []
 mock = []
 stt = ["reqwest/multipart"]
-candle = ["dep:candle-core", "dep:candle-nn", "dep:candle-transformers", "dep:hf-hub", "dep:tokenizers"]
+candle = ["dep:candle-core", "dep:candle-nn", "dep:candle-transformers", "dep:hf-hub", "dep:tokenizers", "dep:symphonia", "dep:rubato"]
 cuda = ["candle", "candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda"]
 metal = ["candle", "candle-core/metal", "candle-nn/metal", "candle-transformers/metal"]
 
@@ -28,6 +28,8 @@ schemars.workspace = true
 serde = { workspace = true, features = ["derive"] }
 serde_json.workspace = true
 tokenizers = { workspace = true, optional = true }
+rubato = { workspace = true, optional = true }
+symphonia = { workspace = true, optional = true }
 tokio = { workspace = true, features = ["rt", "sync", "time"] }
 tokio-stream.workspace = true
 tracing.workspace = true

crates/zeph-llm/README.md

@@ -19,6 +19,7 @@ Defines the `LlmProvider` trait and ships concrete backends for Ollama, Claude,
 | `orchestrator` | Multi-model coordination and fallback |
 | `router` | Model selection and routing logic |
 | `stt` | `SpeechToText` trait and `WhisperProvider` (OpenAI Whisper, feature-gated behind `stt`) |
+| `candle_whisper` | Local offline STT via Candle (whisper-tiny/base/small, feature-gated behind `candle`) |
 | `error` | `LlmError` — unified error type |
 
 **Re-exports:** `LlmProvider`, `LlmError`