Skip to content

Releases

v1.14.2

Choose a tag to compare

@oslatt oslatt released this 21 Nov 17:34
676af8e

Official Fish Audio repository release; no behaviour changes (#148).

  • Official repository: this plugin is now hosted at fishaudio/fish-audio-hermes and maintained by Fish Audio.
  • Updated the install command, plugin.yaml author and homepage, package homepage, issue and security advisory links, and the fish-audio-report-problem skill to point to fishaudio/fish-audio-hermes.
  • Updated the license copyright notice to 39 AI, Inc. GitHub redirects links and git from the previous repository address, so existing installs and links keep working.

Install or update: hermes plugins install fishaudio/fish-audio-hermes --enable --yes-deps

v1.14.1

Choose a tag to compare

@100yenadmin 100yenadmin released this 14 Nov 10:44
8c6bc03

A bug-fix release: Fish's tools now do what an agent asks, or say why not (#127).

  • fish_speak takes output_path. The audio is also saved to the file you name. It never overwrites an existing file, refuses protected locations through Hermes's own write guard, and refuses .., control characters and MEDIA: in the path, all before any billed request. With timestamps, the .srt and .vtt files are copied too. The voice note or file sent to chat is unchanged. On a Hermes without that guard, output_path is refused with a hint to copy from file_path.
  • fish_transcribe takes path as another name for file_path; giving both with different values is refused.
  • Unknown arguments are refused. A Fish tool used to drop an argument it didn't know and report success. Now it names the argument and lists the accepted ones, with no Fish request.
  • README: Fish speech-to-text has no custom-vocabulary option, so choose another provider for jargon-heavy dictation.

Install or update: hermes plugins install 100yenadmin/fish-audio-hermes --enable --yes-deps

Release acceptance passed on Hermes v0.23.1 and on the evaOS runtime at this commit.

v1.14.0

Choose a tag to compare

@100yenadmin 100yenadmin released this 12 Nov 00:24
2c14753

For developers, operators and every language: a scripting CLI, translate-and-voice, voice names in any script, a Desktop voice editor, and operator-mode voices.

  • Scripting CLI: hermes fish speak, transcribe, voices and balance support --json, with safe speech output copies and no terminal prompts (#96).
  • Translate and voice: a new skill transcribes a clip with speaker turns, translates it, and voices it again with one
    Fish voice per speaker, one multi-speaker call per part of up to four speakers. It returns both transcripts and new subtitles, shows the cost
    first, never clones the people in the clip, and calls the result a voice-over, not a synchronized dub (#98).
  • On operator-managed agents, Desktop says which saved name a newly cloned or designed voice got, or why it got none; the fish_voices description says mine lists only saved-name voices there (#122).
  • The Desktop voice editor keeps a detail-load error visible while you type, and the acceptance row reports both errors when restoring a voice title fails after a failed edit (#120).
  • List this profile's saved-name voices in operator mode, automatically name cloned/saved voices, and return billable usage on speech and transcription jobs (#97).
  • Edit your own voice's title and description on Desktop ▸ Voices ▸ My voices (#99).
  • Save voice names in any script, including Japanese, Arabic and Hindi, with Unicode normalization (#99). When two stored names normalize alike, the first in code-point order wins.
  • /fish setup reads the Fish key only through Hermes's secret scope and no longer opens the profile .env; with no key visible it says to restart the gateway or Desktop app after saving one.
  • A reply that mixes a voice clip with a non-voice clip (a design preview, or a WAV without delivery) no longer becomes voice notes; the file wins, as it does for delivery: file (#125).
  • fish_speak with timestamps: true keeps its subtitles when the full alignment file is too large to save, and says so (#125).
  • Desktop: previewing the same voice again while its preview is loading no longer starts a second billed preview; the reading-speed slider follows the saved value unless you have moved it (#125).
  • Long-form skill: the manifest records the output format and subtitle choice, each part's cache path is saved before copying, and an entry whose subtitles failed is kept as audio-only instead of done (#125).

Install or update: hermes plugins install 100yenadmin/fish-audio-hermes --enable --yes-deps

Release acceptance passed on Hermes v0.23.1 and on the evaOS runtime at this commit.

v1.13.0

Choose a tag to compare

@100yenadmin 100yenadmin released this 11 Nov 22:01
0f0c62f

For creators: auditions, voice notes or files, and long-form chapters.

  • Voice auditions: fish_voices audition plays up to four voices reading one short passage, with each billed
    clip attached in chat; the voice-studio skill guides choosing and comparing candidates. Desktop Library and
    My voices get a per-agent "Preview text" box (#93).
  • Voice notes or files: fish_speak accepts delivery: voice (OGG/MP3) or delivery: file (an audio file, any format); a reply that includes a file request is not turned into voice notes.
    Omitting it preserves format-based delivery; WAV voice notes are refused before any Fish request (#95).
  • Reply rules and scheduled briefings: the expressive-speech skill explains Hermes memory/preferences rules,
    and the README includes a morning spoken-briefing cron recipe with a named chat destination (#95).
  • Long-form chapters: a registered skill plans a manuscript, shows the cost estimate before synthesis,
    saves chapter files and a manifest, resumes without redoing finished chapters, and redoes one chapter on request.
    It explains the Story Studio handoff and its separate OAuth sign-in and app credits (#94).
  • Complete word alignment: fish_speak with timestamps: true writes all segments to an .alignment.json
    sidecar and returns alignment_path; the tool result keeps its existing 500-segment cap (#94).
  • Refuse pronunciation dictionary listing, creating, publishing and deletion on operator-managed agents, where dictionaries belong to the operator's whole account (#110).
  • Keep unsaved reading-speed slider changes when adding or removing a saved voice name (#110).
  • Announce the reading-speed slider value as a multiplier to screen readers (#110).
  • Keep dictionary and version IDs up to 128 characters in deletion approvals and rule keys (#110).
  • README: Fish's Live 3 full-duplex voice agents run on Fish's agent platform, not in this plugin; live dictation uses
    transcribe-stream-1.

Install or update: hermes plugins install 100yenadmin/fish-audio-hermes --enable --yes-deps

Release acceptance passed on Hermes v0.23.1 and on the evaOS runtime at this commit.

v1.12.1

Choose a tag to compare

@100yenadmin 100yenadmin released this 09 Nov 22:27
766ed5c

Documentation only; no behaviour changes.

  • New first install line: hermes plugins install 100yenadmin/fish-audio-hermes --enable --yes-deps installs the latest release (on Hermes 0.21.5, leave out --yes-deps). The catalog name comes second: on Hermes 0.23.0 it currently installs an older copy that shipped with Hermes (NousResearch/hermes-agent#151856), and the README says how to replace it.
  • A clearer first screen: what the plugin does in one line, release, license, Hermes and CI badges, and a three-step "Try it" that ends with /fish setup and /fish preview.
  • A settings table for voice, model, reading speed, pronunciation dictionaries, streaming, live dictation, the free-model policy, the API endpoint and operator mode.
  • One troubleshooting section with the bug-report steps, and known limitations now list the Desktop voice-conversation speech gap (NousResearch/hermes-agent#129305) and the faint API key field (NousResearch/hermes-agent#138994).

Release acceptance on the release head (same tree as this tag) passed on Hermes v0.23.0 and the evaOS runtime. Fresh installs on clean Hermes 0.23.0 with no terminal end enabled at 1.12.1.

v1.12.0

Choose a tag to compare

@100yenadmin 100yenadmin released this 09 Nov 13:37
9a93649

Says names right, sets up in chat.

  • Setup you can finish from chat: /fish setup works in any chat before a key is saved. It lists what the profile still needs and the next step: the key (never asked for in chat), a restart when the key was saved after the gateway started, and whether Fish is the voice users actually hear ("Saved for Fish Audio, but this agent currently speaks with edge", with how to switch). It is free. When everything is ready it ends with "Ready. Hear your voice: /fish preview", and /fish preview with no voice now plays a short billed sample in the profile's voice. /fish status and fish_settings show the same next step (#91).
  • Says names right, every time: the free fish_pronunciations tool manages Fish pronunciation dictionaries: create one, publish complete immutable versions, use them as the profile's default for ordinary voice replies, or pass dictionaries to one fish_speak call. Deleting a dictionary or a version goes through Hermes approval (#90, #84).
  • Reading speed in Desktop Settings: a slider in Settings ▸ Plugins ▸ Fish Audio (0.5–2.0×, Save and Reset), and fish_settings set_speed for the agent. Voice replies, read-aloud and streaming use it (#92).

Requires Hermes 0.21.6 or later for live dictation (unchanged).

Release acceptance on the reviewed head (same tree as this tag):

  • Hermes v0.23.0: PASS (95 rows, including the new setup, reading-speed and dictionary rows; expected failures only on open host bugs NousResearch/hermes-agent#138994 and #129305).
  • evaOS runtime: PASS (82 rows).
  • Fresh installs on clean Hermes 0.23.0 with no terminal: hermes plugins install 100yenadmin/fish-audio-hermes --ref <sha> --enable --yes-deps and the URL form end enabled at 1.12.0 with no scan caution.
  • Real chat turns on Hermes 0.23.0: "Save it as Kate" saved a library voice under that name, /fish use Kate switched to it and Hermes spoke with Fish; asked to find a narrator voice and make it its own, the agent did so with its Fish tools.

v1.11.0

Choose a tag to compare

@100yenadmin 100yenadmin released this 09 Nov 11:41
462eab1

The agent runs its own voice.

  • fish_settings: the agent can see and set its own Fish voice, model and saved voice names, with the same rules as /fish (managed settings, operator mode, and never replacing a speech provider that is already set). It also reports which speech provider users actually hear. Ask "find a calm narrator voice and make it your own" and the agent searches, switches and saves the name itself (#87).
  • fish_account: balance and plan (the same read as /fish balance), Fish's models with dated prices and retirement dates, and cost estimates for speech (UTF-8 bytes) or transcription (audio hours) without spending credit. Drama 3's launch price ends on 1 December and estimates follow the date. Operator-managed accounts see no amounts or prices (#88).
  • Replacing an older catalog install on Hermes 0.23.0: hermes plugins remove fish-audio, then the repository install line (README).
  • Maintenance: parity ledger surfaces corrected; the weekly benchmark prices Drama 3 at its list price (#89).

Requires Hermes 0.21.6 or later for live dictation (unchanged).

Release acceptance at the merged tree:

  • Hermes v0.23.0: PASS (88 rows, including the new A-TOOL-SETTINGS and A-TOOL-ACCOUNT rows; expected failures only on open host bugs NousResearch/hermes-agent#138994 and #129305).
  • evaOS runtime: PASS (75 rows).
  • Fresh installs on clean Hermes 0.23.0 with no terminal: the repository URL and owner/repo forms both end enabled at 1.11.0.
  • A real chat turn on Hermes 0.23.0: asked to find a narrator voice and make it its own, the agent called fish_voices search and fish_settings use, add_name and get, with no terminal; the profile's voice and saved name changed as asked.

v1.10.2

Choose a tag to compare

@100yenadmin 100yenadmin released this 09 Nov 11:12
8e15851

Install lines that work when an agent installs the plugin. Documentation only; no behaviour changes.

  • --yes-deps in the README's install commands: on Hermes 0.21.6 and newer, an install with no terminal (an agent installing the plugin for you, or a script) cannot answer the Python-dependency prompt, so Fish Audio ended up installed but disabled. With --enable --yes-deps it installs enabled.
  • Installing on Hermes 0.23.0: the live catalog index is currently over Hermes's 2 MiB client limit (NousResearch/hermes-agent#151856), so a fresh 0.23.0 falls back to the catalog copy it shipped with and the catalog name installs an older release. The README says so and points to the repository line; main only ever holds a tagged, tested release.
  • README corrections: streamed sentences are retried when Fish turns them away before any audio (since 1.7.6), and a key saved in Desktop's Agent settings field works without a restart (since 1.7.5).

Requires Hermes 0.21.6 or later for live dictation (unchanged).

Release acceptance at the merged tree:

  • Hermes v0.23.0: PASS (86 rows; expected failures only on open host bugs NousResearch/hermes-agent#138994 and #129305).
  • evaOS runtime: PASS (73 rows).
  • Fresh installs on clean Hermes 0.23.0 with no terminal: the repository URL and owner/repo forms both end enabled at 1.10.2.

v1.10.1

Choose a tag to compare

@100yenadmin 100yenadmin released this 09 Nov 07:07
298abfa

No stray voice-message marker in Hermes Desktop or the CLI.

  • No [[audio_as_voice]] line under the audio player: after fish_speak, Hermes Desktop and the classic CLI showed the gateway's voice-message marker as a literal text line. On those local surfaces the plugin now adds only the MEDIA: line, and drops a marker the model copied from the tool result. Telegram, Discord and the other messaging gateways get exactly what they got before, so replies there stay voice messages (#83). Desktop shows the same line after core's own text_to_speech; that is reported upstream as NousResearch/hermes-agent#151711.
  • Release acceptance: a run now removes the Fish key from its disposable home when it ends (#82).

Requires Hermes 0.21.6 or later for live dictation (unchanged).

Release acceptance at the merged tree:

v1.10.0

Choose a tag to compare

@100yenadmin 100yenadmin released this 08 Nov 07:22
42a6d3b

Use your own voices by title, and fix pronunciations with plain-text respellings.

  • Use your own voices by title: /fish use Eva, /fish preview Eva and hermes fish use Eva find one of your own Fish voices by its title, ignoring case. Voice ids and saved names still resolve first, with no network call. When several of your voices have that title, or Fish's title search returns more than 100 of your voices, the command lists the choices or asks for the id instead of guessing (#71).

  • Plain-text pronunciation fixes: fish_speak takes a new spoken_as map, such as {"Codex": "code ex"}. Fish reads each replacement as ordinary text (its alias rules), billed on the replaced text. pronunciations still means phonetic forms (CMU Arpabet in English; drama-3 also takes IPA) and sends exactly what it sent before. The expressive-speech skill now explains both. The config's pronunciation_dictionary accepts type: alias per item (#73).

  • Live dictation keeps a combining mark on its letter: when Fish put a sound tag between a letter and its own combining mark, removing the tag left the mark detached (क[x]ि became क ि). It now stays on its letter (कि) (#70).

  • README Disclosure now says:

    • what the report-a-problem and voice-studio skills run, read and send;
    • which commands write the profile's voice settings;
    • that cover images are sent;
    • that a failed live dictation session falls back to file transcription, a separate billed request.

    The /fish argument hint lists alias (#70).

  • Arabic locale: its direction marks are written as escapes in the source, so hermes plugins validate no longer flags invisible Unicode. What ships is unchanged (#69).

  • Release acceptance: the NousResearch/hermes-agent#129305 waivers match only that host bug (#65, #70).

Requires Hermes 0.21.6 or later for live dictation (unchanged).

Release acceptance at the merged tree: