voxcaster 0.2.0__tar.gz

voxcaster-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 ibrahim Alfa
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

voxcaster-0.2.0/PKG-INFO

@@ -0,0 +1,261 @@
+Metadata-Version: 2.4
+Name: voxcaster
+Version: 0.2.0
+Summary: Local text-to-speech daemon and CLI for speech notifications from scripts, builds, cron jobs, and ML training runs, powered by Kokoro TTS with GPU offload and espeak fallback.
+Author: ibrahim Alfa
+License: MIT
+Project-URL: Homepage, https://github.com/I-Alpha/speakd
+Project-URL: Repository, https://github.com/I-Alpha/speakd
+Project-URL: Issues, https://github.com/I-Alpha/speakd/issues
+Project-URL: Changelog, https://github.com/I-Alpha/speakd/blob/master/CHANGELOG.md
+Keywords: tts,text-to-speech,speech-synthesis,speech-notifications,kokoro,kokoro-tts,daemon,unix-socket,cli,narration,notifications,machine-learning,gpu,cuda,developer-tools
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: No Input/Output (Daemon)
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Operating System :: MacOS
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
+Classifier: Topic :: System :: Monitoring
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: kokoro>=0.9
+Requires-Dist: soundfile>=0.12
+Requires-Dist: numpy>=1.24
+Requires-Dist: tomli>=2.0; python_version < "3.11"
+Dynamic: license-file
+
+# speakd
+
+[![PyPI](https://img.shields.io/pypi/v/speakd)](https://pypi.org/project/speakd/)
+[![PyPI - Python Version](https://img.shields.io/pypi/pyversions/speakd)](https://pypi.org/project/speakd/)
+
+**A local text-to-speech daemon and CLI for speech notifications from
+long-running jobs.**
+
+`speakd` is a Python TTS daemon powered by
+[Kokoro](https://github.com/hexgrad/kokoro) (a fast, high-quality local
+text-to-speech model). Shell scripts, machine-learning training runs, builds,
+cron jobs, CI hooks, and Python programs can send fire-and-forget speech
+notifications over a Unix socket; the caller returns in about a millisecond
+while the daemon queues, synthesizes, and plays each line in order. If anything
+in the audio stack fails, the line degrades to espeak instead of disappearing.
+
+It was built to narrate machine-learning training runs on a single-GPU
+workstation, which shaped its defining feature: **the TTS model dynamically
+offloads itself from the GPU** when narration goes quiet, so it never holds
+VRAM hostage from the workload it is narrating.
+
+```
+$ pip install speakd
+$ speak "training started"          # daemon auto-spawns on first use
+$ speak --interrupt "loss is NaN"   # cuts off whatever is playing, speaks NOW
+$ make 2>&1 | tail -1 | speak       # pipe-friendly
+```
+
+## Use cases
+
+- Add voice alerts to machine-learning training runs when epochs finish,
+  checkpoints save, loss becomes NaN, or jobs crash.
+- Turn shell scripts, Makefiles, cron jobs, and CI hooks into spoken status
+  updates.
+- Use Kokoro TTS locally without blocking the process that asked for speech.
+- Share one text-to-speech queue across multiple processes so messages do not
+  overlap.
+- Release GPU VRAM after narration bursts with dynamic CPU/GPU offload.
+
+## Why a daemon?
+
+Calling a TTS library inline is the obvious approach and the wrong one for
+narration: it blocks the caller for seconds per line, loads a model per
+process, and overlapping lines talk over each other. `speakd` inverts this:
+
+- **~1 ms per call.** The client writes one line to a Unix socket and returns.
+  Narration can sit inside hot loops and signal handlers.
+- **One model, one queue.** A single daemon owns the model and serialises
+  playback. Ten processes can narrate concurrently without crosstalk.
+- **Failure-proof by design.** Daemon down? The client spawns it. Spawn fails?
+  espeak fallback. No audio at all? The caller still never raises.
+
+## Architecture
+
+```
+ any process, any language                 speakd daemon (one per socket, flock-enforced)
+┌──────────────────────┐            ┌───────────────────────────────────────────────┐
+│  speak "epoch done"  │──┐         │   asyncio Unix-socket server                  │
+└──────────────────────┘  │         │        │                                      │
+┌──────────────────────┐  │  UTF-8  │        ├── volume msg ──▶ live volume         │
+│  Python: speak(...)  │──┼─ line ─▶│        ├── interrupt ───▶ drain queue +       │
+└──────────────────────┘  │  over   │        │                  kill playback       │
+┌──────────────────────┐  │  socket │        ▼                                      │
+│  CI job, cron, hook  │──┘         │   FIFO queue ──▶ worker (thread executor)     │
+└──────────────────────┘            │                     │                         │
+          ▲                         │                     ▼                         │
+          │ "OK\n" ack              │   Kokoro TTS ──▶ wav ──▶ mpv ──▶ 🔊           │
+          │ (blocking mode only)    │   CPU ⇄ GPU                                   │
+          └─────────────────────────│   (offloads after idle keepalive)             │
+                                    │                                               │
+                                    │   any failure ──▶ espeak fallback             │
+                                    └───────────────────────────────────────────────┘
+```
+
+## Features
+
+- **Fire-and-forget socket design** — newline-terminated UTF-8 over a Unix
+  domain socket; trivially scriptable from any language. Optional `OK` ack
+  for blocking callers.
+- **Dynamic GPU offload with keepalive** — the model loads on CPU, hops onto
+  the GPU for narration bursts, and releases its VRAM (~3 GB) after a
+  configurable idle period. If the GPU is full (another job grabbed it), that
+  request simply synthesizes on CPU instead of failing.
+- **Interrupt protocol** — an urgent line drains the pending queue, kills
+  in-flight playback mid-word, and speaks immediately.
+- **Live volume control** — one socket message, applies from the next line;
+  no restart.
+- **Singleton via `flock(2)`** — clients can race to auto-spawn the daemon;
+  exactly one wins, the rest exit cleanly. Stale sockets are detected and
+  removed on startup.
+- **Graceful fallback** — Kokoro import error, synthesis failure, playback
+  failure, or daemon unreachable: the line is spoken by espeak and the event
+  is logged. Narration degrades; it never silently vanishes.
+- **One TOML file, env-var overrides, zero-config defaults** — works out of
+  the box on CPU with no config file at all.
+
+## Requirements
+
+- Linux or macOS (Unix sockets + `flock`), Python ≥ 3.10
+- [mpv](https://mpv.io/) for playback (`apt install mpv`) — or any player,
+  via config
+- [espeak](https://espeak.sourceforge.net/) for the fallback voice
+  (`apt install espeak`) — optional but recommended
+- A CUDA-capable GPU is **optional**; everything works on CPU
+
+## Install
+
+```bash
+pip install speakd
+```
+
+This installs the `kokoro` TTS package (which pulls in PyTorch) and two
+console commands: `speakd` (the daemon) and `speak` (the client).
+
+To install from source:
+
+```bash
+git clone https://github.com/I-Alpha/speakd && cd speakd
+pip install .
+```
+
+## Quickstart
+
+```bash
+# 1. Just speak — the daemon auto-spawns on first use:
+speak "hello from speakd"
+
+# 2. Or run the daemon in the foreground to watch it work:
+speakd --device cpu --voice af_heart
+
+# 3. Script it:
+speak --blocking "waits until this has been spoken"
+speak --interrupt "queue drained, this plays immediately"
+speak --volume 60 "quieter from now on"
+echo "pipes work too" | speak
+```
+
+From Python:
+
+```python
+from speakd import speak, set_volume
+
+speak("checkpoint saved")                        # ~1 ms, non-blocking
+speak("eval finished", blocking=True)            # wait until spoken
+speak("loss is NaN — stopping", interrupt=True)  # jump the queue
+set_volume(85)
+```
+
+See [`examples/`](examples/) for runnable demos of narration, interrupts,
+and volume control.
+
+## Configuration
+
+Defaults work with no config at all. To customise, copy
+[`config.example.toml`](config.example.toml) to `~/.config/speakd/config.toml`
+(or point `$SPEAKD_CONFIG` at any path). Environment variables override the
+file; CLI flags override both.
+
+| TOML key | Env override | Default | Meaning |
+|---|---|---|---|
+| `tts.voice` | `SPEAKD_VOICE` | `af_heart` | Kokoro voice id (`af_*`, `am_*`, `bf_*`, `bm_*`, ...) |
+| `tts.speed` | `SPEAKD_SPEED` | `1.0` | Speech-rate multiplier |
+| `tts.lang_code` | `SPEAKD_LANG` | `a` | Kokoro language code (`a` US English, `b` UK English) |
+| `device.policy` | `SPEAKD_DEVICE` | `auto` | `auto` (dynamic offload) / `cpu` / `gpu` |
+| `device.keepalive_seconds` | `SPEAKD_KEEPALIVE` | `180` | Idle seconds before GPU→CPU offload |
+| `daemon.socket_path` | `SPEAKD_SOCKET` | `$XDG_RUNTIME_DIR/speakd.sock` | Unix socket path |
+| `daemon.socket_mode` | — | `"600"` | Octal permissions on the socket file |
+| `daemon.log_file` | `SPEAKD_LOG_FILE` | `~/.local/state/speakd/daemon.log` | Log target for auto-spawned daemons |
+| `audio.volume` | `SPEAKD_VOLUME` | `100` | Playback volume `0–130` (mpv scale) |
+| `audio.max_playback_seconds` | — | `120` | Kill a single line's playback after this |
+| `audio.player` | — | mpv template | Player argv; `{file}` and `{volume}` are substituted |
+| `fallback.command` | — | espeak template | Fallback argv; `{text}` is substituted; `[]` disables |
+| `client.connect_timeout` | — | `0.5` | Socket connect/send timeout (s) |
+| `client.ack_timeout` | — | `300.0` | `--blocking` wait for the spoken-ack (s) |
+| `client.spawn_wait` | — | `4.0` | Wait for an auto-spawned daemon (s) |
+
+`speakd --print-config` shows the fully-resolved effective configuration.
+
+## Wire protocol
+
+One newline-terminated UTF-8 line per connection — easy to speak from any
+language without a client library:
+
+| Message | Bytes | Effect |
+|---|---|---|
+| Speak | `<text>\n` | Queue the line; daemon replies `OK\n` when spoken |
+| Interrupt | `\x01INTERRUPT\x01<text>\n` | Drain queue, kill playback, speak now |
+| Volume | `\x02VOLUME\x02<int>\n` | Set live volume (0–130) |
+
+```bash
+# speak from raw shell, no client needed:
+printf 'hello from netcat\n' | nc -U "$XDG_RUNTIME_DIR/speakd.sock"
+```
+
+The control markers are ASCII SOH/STX characters that cannot occur in normal
+text, so no escaping is ever needed.
+
+## GPU offload in detail
+
+The `auto` policy exists for machines where the GPU has a day job:
+
+1. The model loads on **CPU** at first request.
+2. Each synthesis tries to move it to the **GPU** first (a few hundred ms,
+   then synthesis is much faster). If CUDA is busy or OOM, that line
+   synthesizes on CPU — no error, just slower.
+3. After `keepalive_seconds` (default 180 s) without a request, an idle timer
+   moves the model back to **CPU** and calls `torch.cuda.empty_cache()`,
+   releasing the VRAM.
+
+The effect: during an active narration burst the voice is snappy and
+GPU-accelerated; ten minutes into a silent stretch, your training job has its
+VRAM back. All device moves are serialised with synthesis under one lock, so
+the model can never be moved mid-utterance.
+
+## Troubleshooting
+
+| Symptom | Likely cause / fix |
+|---|---|
+| `speak` says *fallback engine used* | Daemon failed to start — check `~/.local/state/speakd/daemon.log`. Most common: `kokoro` not installed in the Python that spawned it (set `SPEAKD_DAEMON_CMD="/path/to/python -m speakd.daemon"`). |
+| No audio, no errors | Is `mpv` installed and does it play a wav from your terminal? Swap `audio.player` if you use a different player. |
+| First line is slow | Cold start: model weights load on first request (a few seconds). Subsequent lines are fast. |
+| Robotic voice instead of Kokoro | That *is* the espeak fallback working as designed — see the first row. |
+| Two daemons after a crash | They cannot coexist: the flock singleton makes the second exit immediately, and stale sockets are cleaned on startup. Delete `<socket>.lock` only if a machine crash left it owned by a dead PID holder (flock releases on process death, so this is near-impossible). |
+| `daemon already running (pid N)` | Working as intended — the running daemon serves all clients. |
+| GPU memory not released | The model offloads after `device.keepalive_seconds` of *no requests*; lower it, or run with `--device cpu`. |
+
+## License
+
+[MIT](LICENSE) © 2026 ibrahim Alfa

voxcaster-0.2.0/README.md

@@ -0,0 +1,229 @@
+# speakd
+
+[![PyPI](https://img.shields.io/pypi/v/speakd)](https://pypi.org/project/speakd/)
+[![PyPI - Python Version](https://img.shields.io/pypi/pyversions/speakd)](https://pypi.org/project/speakd/)
+
+**A local text-to-speech daemon and CLI for speech notifications from
+long-running jobs.**
+
+`speakd` is a Python TTS daemon powered by
+[Kokoro](https://github.com/hexgrad/kokoro) (a fast, high-quality local
+text-to-speech model). Shell scripts, machine-learning training runs, builds,
+cron jobs, CI hooks, and Python programs can send fire-and-forget speech
+notifications over a Unix socket; the caller returns in about a millisecond
+while the daemon queues, synthesizes, and plays each line in order. If anything
+in the audio stack fails, the line degrades to espeak instead of disappearing.
+
+It was built to narrate machine-learning training runs on a single-GPU
+workstation, which shaped its defining feature: **the TTS model dynamically
+offloads itself from the GPU** when narration goes quiet, so it never holds
+VRAM hostage from the workload it is narrating.
+
+```
+$ pip install speakd
+$ speak "training started"          # daemon auto-spawns on first use
+$ speak --interrupt "loss is NaN"   # cuts off whatever is playing, speaks NOW
+$ make 2>&1 | tail -1 | speak       # pipe-friendly
+```
+
+## Use cases
+
+- Add voice alerts to machine-learning training runs when epochs finish,
+  checkpoints save, loss becomes NaN, or jobs crash.
+- Turn shell scripts, Makefiles, cron jobs, and CI hooks into spoken status
+  updates.
+- Use Kokoro TTS locally without blocking the process that asked for speech.
+- Share one text-to-speech queue across multiple processes so messages do not
+  overlap.
+- Release GPU VRAM after narration bursts with dynamic CPU/GPU offload.
+
+## Why a daemon?
+
+Calling a TTS library inline is the obvious approach and the wrong one for
+narration: it blocks the caller for seconds per line, loads a model per
+process, and overlapping lines talk over each other. `speakd` inverts this:
+
+- **~1 ms per call.** The client writes one line to a Unix socket and returns.
+  Narration can sit inside hot loops and signal handlers.
+- **One model, one queue.** A single daemon owns the model and serialises
+  playback. Ten processes can narrate concurrently without crosstalk.
+- **Failure-proof by design.** Daemon down? The client spawns it. Spawn fails?
+  espeak fallback. No audio at all? The caller still never raises.
+
+## Architecture
+
+```
+ any process, any language                 speakd daemon (one per socket, flock-enforced)
+┌──────────────────────┐            ┌───────────────────────────────────────────────┐
+│  speak "epoch done"  │──┐         │   asyncio Unix-socket server                  │
+└──────────────────────┘  │         │        │                                      │
+┌──────────────────────┐  │  UTF-8  │        ├── volume msg ──▶ live volume         │
+│  Python: speak(...)  │──┼─ line ─▶│        ├── interrupt ───▶ drain queue +       │
+└──────────────────────┘  │  over   │        │                  kill playback       │
+┌──────────────────────┐  │  socket │        ▼                                      │
+│  CI job, cron, hook  │──┘         │   FIFO queue ──▶ worker (thread executor)     │
+└──────────────────────┘            │                     │                         │
+          ▲                         │                     ▼                         │
+          │ "OK\n" ack              │   Kokoro TTS ──▶ wav ──▶ mpv ──▶ 🔊           │
+          │ (blocking mode only)    │   CPU ⇄ GPU                                   │
+          └─────────────────────────│   (offloads after idle keepalive)             │
+                                    │                                               │
+                                    │   any failure ──▶ espeak fallback             │
+                                    └───────────────────────────────────────────────┘
+```
+
+## Features
+
+- **Fire-and-forget socket design** — newline-terminated UTF-8 over a Unix
+  domain socket; trivially scriptable from any language. Optional `OK` ack
+  for blocking callers.
+- **Dynamic GPU offload with keepalive** — the model loads on CPU, hops onto
+  the GPU for narration bursts, and releases its VRAM (~3 GB) after a
+  configurable idle period. If the GPU is full (another job grabbed it), that
+  request simply synthesizes on CPU instead of failing.
+- **Interrupt protocol** — an urgent line drains the pending queue, kills
+  in-flight playback mid-word, and speaks immediately.
+- **Live volume control** — one socket message, applies from the next line;
+  no restart.
+- **Singleton via `flock(2)`** — clients can race to auto-spawn the daemon;
+  exactly one wins, the rest exit cleanly. Stale sockets are detected and
+  removed on startup.
+- **Graceful fallback** — Kokoro import error, synthesis failure, playback
+  failure, or daemon unreachable: the line is spoken by espeak and the event
+  is logged. Narration degrades; it never silently vanishes.
+- **One TOML file, env-var overrides, zero-config defaults** — works out of
+  the box on CPU with no config file at all.
+
+## Requirements
+
+- Linux or macOS (Unix sockets + `flock`), Python ≥ 3.10
+- [mpv](https://mpv.io/) for playback (`apt install mpv`) — or any player,
+  via config
+- [espeak](https://espeak.sourceforge.net/) for the fallback voice
+  (`apt install espeak`) — optional but recommended
+- A CUDA-capable GPU is **optional**; everything works on CPU
+
+## Install
+
+```bash
+pip install speakd
+```
+
+This installs the `kokoro` TTS package (which pulls in PyTorch) and two
+console commands: `speakd` (the daemon) and `speak` (the client).
+
+To install from source:
+
+```bash
+git clone https://github.com/I-Alpha/speakd && cd speakd
+pip install .
+```
+
+## Quickstart
+
+```bash
+# 1. Just speak — the daemon auto-spawns on first use:
+speak "hello from speakd"
+
+# 2. Or run the daemon in the foreground to watch it work:
+speakd --device cpu --voice af_heart
+
+# 3. Script it:
+speak --blocking "waits until this has been spoken"
+speak --interrupt "queue drained, this plays immediately"
+speak --volume 60 "quieter from now on"
+echo "pipes work too" | speak
+```
+
+From Python:
+
+```python
+from speakd import speak, set_volume
+
+speak("checkpoint saved")                        # ~1 ms, non-blocking
+speak("eval finished", blocking=True)            # wait until spoken
+speak("loss is NaN — stopping", interrupt=True)  # jump the queue
+set_volume(85)
+```
+
+See [`examples/`](examples/) for runnable demos of narration, interrupts,
+and volume control.
+
+## Configuration
+
+Defaults work with no config at all. To customise, copy
+[`config.example.toml`](config.example.toml) to `~/.config/speakd/config.toml`
+(or point `$SPEAKD_CONFIG` at any path). Environment variables override the
+file; CLI flags override both.
+
+| TOML key | Env override | Default | Meaning |
+|---|---|---|---|
+| `tts.voice` | `SPEAKD_VOICE` | `af_heart` | Kokoro voice id (`af_*`, `am_*`, `bf_*`, `bm_*`, ...) |
+| `tts.speed` | `SPEAKD_SPEED` | `1.0` | Speech-rate multiplier |
+| `tts.lang_code` | `SPEAKD_LANG` | `a` | Kokoro language code (`a` US English, `b` UK English) |
+| `device.policy` | `SPEAKD_DEVICE` | `auto` | `auto` (dynamic offload) / `cpu` / `gpu` |
+| `device.keepalive_seconds` | `SPEAKD_KEEPALIVE` | `180` | Idle seconds before GPU→CPU offload |
+| `daemon.socket_path` | `SPEAKD_SOCKET` | `$XDG_RUNTIME_DIR/speakd.sock` | Unix socket path |
+| `daemon.socket_mode` | — | `"600"` | Octal permissions on the socket file |
+| `daemon.log_file` | `SPEAKD_LOG_FILE` | `~/.local/state/speakd/daemon.log` | Log target for auto-spawned daemons |
+| `audio.volume` | `SPEAKD_VOLUME` | `100` | Playback volume `0–130` (mpv scale) |
+| `audio.max_playback_seconds` | — | `120` | Kill a single line's playback after this |
+| `audio.player` | — | mpv template | Player argv; `{file}` and `{volume}` are substituted |
+| `fallback.command` | — | espeak template | Fallback argv; `{text}` is substituted; `[]` disables |
+| `client.connect_timeout` | — | `0.5` | Socket connect/send timeout (s) |
+| `client.ack_timeout` | — | `300.0` | `--blocking` wait for the spoken-ack (s) |
+| `client.spawn_wait` | — | `4.0` | Wait for an auto-spawned daemon (s) |
+
+`speakd --print-config` shows the fully-resolved effective configuration.
+
+## Wire protocol
+
+One newline-terminated UTF-8 line per connection — easy to speak from any
+language without a client library:
+
+| Message | Bytes | Effect |
+|---|---|---|
+| Speak | `<text>\n` | Queue the line; daemon replies `OK\n` when spoken |
+| Interrupt | `\x01INTERRUPT\x01<text>\n` | Drain queue, kill playback, speak now |
+| Volume | `\x02VOLUME\x02<int>\n` | Set live volume (0–130) |
+
+```bash
+# speak from raw shell, no client needed:
+printf 'hello from netcat\n' | nc -U "$XDG_RUNTIME_DIR/speakd.sock"
+```
+
+The control markers are ASCII SOH/STX characters that cannot occur in normal
+text, so no escaping is ever needed.
+
+## GPU offload in detail
+
+The `auto` policy exists for machines where the GPU has a day job:
+
+1. The model loads on **CPU** at first request.
+2. Each synthesis tries to move it to the **GPU** first (a few hundred ms,
+   then synthesis is much faster). If CUDA is busy or OOM, that line
+   synthesizes on CPU — no error, just slower.
+3. After `keepalive_seconds` (default 180 s) without a request, an idle timer
+   moves the model back to **CPU** and calls `torch.cuda.empty_cache()`,
+   releasing the VRAM.
+
+The effect: during an active narration burst the voice is snappy and
+GPU-accelerated; ten minutes into a silent stretch, your training job has its
+VRAM back. All device moves are serialised with synthesis under one lock, so
+the model can never be moved mid-utterance.
+
+## Troubleshooting
+
+| Symptom | Likely cause / fix |
+|---|---|
+| `speak` says *fallback engine used* | Daemon failed to start — check `~/.local/state/speakd/daemon.log`. Most common: `kokoro` not installed in the Python that spawned it (set `SPEAKD_DAEMON_CMD="/path/to/python -m speakd.daemon"`). |
+| No audio, no errors | Is `mpv` installed and does it play a wav from your terminal? Swap `audio.player` if you use a different player. |
+| First line is slow | Cold start: model weights load on first request (a few seconds). Subsequent lines are fast. |
+| Robotic voice instead of Kokoro | That *is* the espeak fallback working as designed — see the first row. |
+| Two daemons after a crash | They cannot coexist: the flock singleton makes the second exit immediately, and stale sockets are cleaned on startup. Delete `<socket>.lock` only if a machine crash left it owned by a dead PID holder (flock releases on process death, so this is near-impossible). |
+| `daemon already running (pid N)` | Working as intended — the running daemon serves all clients. |
+| GPU memory not released | The model offloads after `device.keepalive_seconds` of *no requests*; lower it, or run with `--device cpu`. |
+
+## License
+
+[MIT](LICENSE) © 2026 ibrahim Alfa

voxcaster-0.2.0/pyproject.toml

@@ -0,0 +1,62 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "voxcaster"
+version = "0.2.0"
+description = "Local text-to-speech daemon and CLI for speech notifications from scripts, builds, cron jobs, and ML training runs, powered by Kokoro TTS with GPU offload and espeak fallback."
+readme = "README.md"
+license = { text = "MIT" }
+authors = [{ name = "ibrahim Alfa" }]
+requires-python = ">=3.10"
+keywords = [
+    "tts",
+    "text-to-speech",
+    "speech-synthesis",
+    "speech-notifications",
+    "kokoro",
+    "kokoro-tts",
+    "daemon",
+    "unix-socket",
+    "cli",
+    "narration",
+    "notifications",
+    "machine-learning",
+    "gpu",
+    "cuda",
+    "developer-tools",
+]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Environment :: No Input/Output (Daemon)",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: POSIX :: Linux",
+    "Operating System :: MacOS",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Multimedia :: Sound/Audio :: Speech",
+    "Topic :: System :: Monitoring",
+]
+dependencies = [
+    "kokoro>=0.9",
+    "soundfile>=0.12",
+    "numpy>=1.24",
+    "tomli>=2.0; python_version < '3.11'",
+]
+
+[project.urls]
+Homepage = "https://github.com/I-Alpha/speakd"
+Repository = "https://github.com/I-Alpha/speakd"
+Issues = "https://github.com/I-Alpha/speakd/issues"
+Changelog = "https://github.com/I-Alpha/speakd/blob/master/CHANGELOG.md"
+
+[project.scripts]
+voxcaster = "voxcaster.daemon:main"
+speak = "voxcaster.client:main"
+
+[tool.setuptools.packages.find]
+where = ["src"]

voxcaster-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

voxcaster-0.2.0/src/speakd/__init__.py

@@ -0,0 +1,47 @@
+"""speakd — fire-and-forget local TTS narration over a Unix socket.
+
+A small daemon that turns text lines into speech with `Kokoro
+<https://github.com/hexgrad/kokoro>`_, plus a zero-dependency client.
+Designed for narrating long-running work (training runs, builds, pipelines)
+without ever blocking or crashing the thing doing the work.
+
+Quickstart::
+
+    from speakd import speak
+    speak("experiment finished")                 # fire-and-forget
+    speak("loss is NaN — stopping", interrupt=True)
+"""
+from typing import TYPE_CHECKING
+
+__version__ = "0.1.0"
+
+if TYPE_CHECKING:  # real imports for type checkers / IDEs
+    from .client import ensure_daemon, ping, set_volume, speak
+    from .config import Config, load_config
+    from .markdown import extract_tts_summary, preprocess_for_speech, strip_markdown
+
+__all__ = [
+    "speak", "ping", "set_volume", "ensure_daemon",
+    "Config", "load_config",
+    "strip_markdown", "extract_tts_summary", "preprocess_for_speech",
+    "__version__",
+]
+
+_CLIENT_ATTRS = ("speak", "ping", "set_volume", "ensure_daemon")
+_CONFIG_ATTRS = ("Config", "load_config")
+_MARKDOWN_ATTRS = ("strip_markdown", "extract_tts_summary", "preprocess_for_speech")
+
+
+def __getattr__(name: str):
+    """Lazy re-exports (PEP 562): keep ``import speakd`` instant and avoid
+    eagerly importing submodules that ``python -m speakd.<mod>`` re-executes."""
+    if name in _CLIENT_ATTRS:
+        from . import client
+        return getattr(client, name)
+    if name in _CONFIG_ATTRS:
+        from . import config
+        return getattr(config, name)
+    if name in _MARKDOWN_ATTRS:
+        from . import markdown
+        return getattr(markdown, name)
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")