Short:        Neural speech narrator.device via Piper
Author:       simond@irrelevant.org (Simon Dick)
Uploader:     simond irrelevant org (Simon Dick)
Type:         util/sys
Version:      44.0
Architecture: m68k-amigaos
Distribution: Aminet

narrator.wyoming
================

A drop-in replacement for the Amiga narrator.device that speaks with a
modern neural voice instead of the old Paula formant synthesizer. Rather
than running local formant synthesis, it forwards text to a Piper neural
text-to-speech server over the Wyoming protocol (plain TCP) and plays the
returned PCM audio through AHI.

Because it replaces narrator.device by name in DEVS:, existing Amiga
software gets neural speech transparently - the stock Say command, and any
program that opens narrator.device, work with no modification.

It ships as a pair:

  * narrator.device    - the drop-in device (goes in DEVS:)
  * translator.library - a pass-through translator so callers deliver plain
                         English to the device (Piper wants English, not the
                         classic ARPABET phonemes)

Aimed at PiStorm / emu68k accelerated and other fast 68k Amigas, where the
accelerated CPU makes the network round-trip practical. Developed and tested
under the Amiberry emulator.


REQUIREMENTS
------------

  * AHI v4 or later, with a Unit 0 audio mode configured (the paula.audio
    driver is fine under emulation).
  * A TCP/IP stack providing bsdsocket.library (Roadshow or AmiTCP).
  * A fast 68k (PiStorm/emu68k, or roughly 68030/40MHz and up). No FPU
    required.
  * A Piper TTS server reachable on your LAN with the Wyoming protocol
    enabled (default port 10200), e.g. the wyoming-piper add-on/container.
  * No TLS / AmiSSL - Wyoming is plain TCP.


INSTALLATION
------------

  Copy narrator.device     to DEVS:narrator.device
  Copy translator.library  to LIBS:translator.library

  Create the prefs file ENV:narrator.wyoming (and ENVARC: to persist it
  across reboots) with at least your server address:

      host 192.168.1.50
      port 10200

  Then use speech as normal, for example:

      Say "Hello from the Amiga."


CONFIGURATION (ENV:narrator.wyoming)
------------------------------------

  "key value" per line; # or ; start comments. All keys are optional
  except host.

    host          Piper/Wyoming server address        (default 127.0.0.1)
    port          server port                         (default 10200)
    voice         default Piper voice name             (server default)
    voice_male    voice when the caller selects MALE   (Say's default sex)
    voice_female  voice when the caller selects FEMALE
    ahi_unit      ahi.device unit to play through      (default 0)
    split_words   break long input into ~N-word
                  pipelined chunks for a faster start  (0 = off)

  Voice names must exist on your Piper server. Choose a clearly-articulating
  voice for voice_male - it is the everyday voice, since Say defaults to the
  MALE sex.


COMPANION SOFTWARE
------------------

  Anything that speaks through narrator.device benefits from the neural
  voice. A good companion is speak-handler by Alexander Fritsch (Aminet:
  util/sys/speak-handler), a from-scratch native replacement for the
  Commodore SPEAK: handler. Mount its SPEAK: and you can pipe text straight
  to neural speech:

      Type myfile.txt TO SPEAK:
      echo "Hello from the Amiga" >SPEAK:


NOTES / LIMITATIONS
-------------------

  * Latency: warm first-audio is around 0.4s on an emulated 68020 over a
    local network. The first request after the server has been idle is
    slower (~1.8s) while Piper loads the voice model; warm requests are
    fast. split_words cuts the start delay on long sentences.

  * rate and pitch from the IOSpeech request are accepted but ignored - the
    Wyoming synthesize request has no per-request rate/pitch knob (those are
    properties of the Piper voice, set server-side). volume and sex are
    honoured (sex selects the configured voice).

  * Direct ARPABET phoneme input (the classic narrator contract, bypassing
    translator.library) is silently discarded - Piper takes text, not
    phonemes. Normal English, including all-caps words, is always spoken.

  * CMD_READ mouth-shapes and word/syllable sync are not implemented (Piper
    returns no phoneme timing). CMD_READ returns ND_NoWrite, so talking-head
    software gets no animation but does not hang.


DEVELOPMENT
-----------

  This project was developed with the assistance of AI tooling (Anthropic's
  Claude, via Claude Code), under human direction and tested on-target.


SOURCE
------

  Full source, build instructions and design notes:

      https://github.com/sidick/narrator.wyoming

  Built with the Bebbo m68k-amigaos GCC cross-toolchain.


LICENSE
-------

  MIT - see the LICENSE file in the source repository.