shinshin86/connect-colab-local-tts

Connect Colab Local TTS

Goal

Use Colab MCP Go to drive Google Colab, launch shinshin86/local-tts-on-google-colab, capture its public trycloudflare URL, and verify speech generation through @aituber-onair/voice.

This skill consumes an existing OpenAI-compatible Colab TTS project. Do not rewrite TTS wrappers unless the user explicitly asks for wrapper development.

Inputs

Collect or infer:

• repo_url: default https://github.com/shinshin86/local-tts-on-google-colab.git
• repo_ref: default main; prefer a tag or commit for reproducible runs
• engine: default Irodori-TTS when the user does not choose one
• model: OpenAI-compatible model id; default to the selected engine
• voice: optional; leave empty for engines with no useful voice selector
• text: Japanese smoke text to synthesize

Good sample engines:

• Irodori-TTS: useful Japanese default candidate
• Piper-Plus: lightweight multilingual candidate
• MOSS-TTS-v1.5: notable Hugging Face model, but A100-class Colab is usually required

Procedure

1. Confirm Colab MCP Go is available.
   • Use open_colab_browser_connection when the Colab session is not connected.
   • Then call list_colab_tools.
   • Do not assume exact remote tool names; choose the available tool that can create/replace notebook cell code and run it.
2. If the run needs a GPU, gated model access, or HF_TOKEN, ask the user to prepare those items in Colab before launching.
   • The user handles Google login, runtime selection, gated-model license acceptance, and token/secret setup.
   • The agent should still explain the concrete missing prerequisite.
3. Generate one Colab cell that launches the selected engine.
   • Use repo_url, repo_ref, and engine.
   • Keep EXPOSE_PUBLIC_URL = True.
   • Pass engine-specific options only when the user chose them.
   • Set OPENAI_MODEL_ID to model when provided.
4. Execute the cell through the Colab MCP tools and monitor output.
   • Wait for === Public Ready ===.
   • Extract Speech : https://...trycloudflare.com/v1/audio/speech.
   • If the MCP result only shows CompletedProcess(...), inspect /content/openai-compatible-local-tts/logs/cloudflared.log and extract the latest https://...trycloudflare.com URL.
   • If the run fails, inspect the tail shown by the Colab output before changing settings.
5. Validate the Colab server itself.
   • Use the generated test curl from Colab output when possible.
   • Confirm GET /v1/models and POST /v1/audio/speech work before running the local voice package example.
6. Validate @aituber-onair/voice locally.
   • Build the voice package before running the Node example.
   • Run packages/voice/examples/node-basic/openai-compatible-colab-example.js with dynamic environment variables:

OPENAI_COMPATIBLE_TTS_URL="<speech_url>" \
OPENAI_COMPATIBLE_TTS_MODEL="<model>" \
OPENAI_COMPATIBLE_TTS_VOICE="<voice>" \
OPENAI_COMPATIBLE_TTS_TEXT="<text>" \
node packages/voice/examples/node-basic/openai-compatible-colab-example.js

Use OPENAI_COMPATIBLE_TTS_PLAY=0 when the environment cannot play audio but WAV generation should still be verified.

Colab Cell Shape

Use the current canonical launch flow from local-tts-on-google-colab. The generated cell should be equivalent to:

REPO_URL = "https://github.com/shinshin86/local-tts-on-google-colab.git"
REPO_REF = "main"
WORKDIR = "/content/local-tts-on-google-colab"
ENGINE = "Irodori-TTS"
EXPOSE_PUBLIC_URL = True
TEST_TEXT = "こんにちは。AITuber OnAir Voice から再生しています。"
TEST_SPEED = 1.0
TEST_VOICE = ""
OPENAI_MODEL_ID = "Irodori-TTS"

import os
import subprocess

if not os.path.exists(WORKDIR):
    subprocess.run(["git", "clone", REPO_URL, WORKDIR], check=True)

subprocess.run(["git", "fetch", "--all", "--tags"], cwd=WORKDIR, check=True)
subprocess.run(["git", "checkout", REPO_REF], cwd=WORKDIR, check=True)

cmd = [
    "python", "colab/bootstrap.py",
    "--engine", ENGINE,
    "--test-text", TEST_TEXT,
    "--test-speed", str(TEST_SPEED),
    "--openai-model-id", OPENAI_MODEL_ID,
]
if TEST_VOICE:
    cmd.extend(["--test-voice", TEST_VOICE])
if EXPOSE_PUBLIC_URL:
    cmd.append("--expose-public-url")
else:
    cmd.append("--no-expose-public-url")

subprocess.run(cmd, cwd=WORKDIR, check=True)

For Piper-Plus, add --piper-plus-model <value> only when the user chooses one. For MOSS-TTS-v1.5, warn that A100 is expected and pass --moss-tts-v1-5-language, --moss-tts-v1-5-default-voice, or prompt WAV options only when selected.

Verification Commands

Run from the repository root after a code change:

npm -w @aituber-onair/voice run fmt
npm -w @aituber-onair/voice run build
npm -w @aituber-onair/voice run test

For a live Colab endpoint:

OPENAI_COMPATIBLE_TTS_URL="<speech_url>" \
OPENAI_COMPATIBLE_TTS_MODEL="<model>" \
OPENAI_COMPATIBLE_TTS_VOICE="<voice>" \
OPENAI_COMPATIBLE_TTS_PLAY=0 \
node packages/voice/examples/node-basic/openai-compatible-colab-example.js

Failure Modes

• Colab MCP tools are unavailable: open the Colab browser connection and list tools again.
• No public URL appears: check cloudflared output and whether the server became ready locally.
• Colab reports success but no detailed output: read the files under /content/openai-compatible-local-tts/logs, especially cloudflared.log and the engine-specific uvicorn log.
• MOSS-TTS-v1.5 fails on L4/T4: retry on A100 or choose a lighter engine.
• POST /v1/audio/speech fails but /v1/models works: inspect Colab logs and verify model, voice, prompt WAV, language, and gated-model prerequisites.
• Node example saves no audio: confirm OPENAI_COMPATIBLE_TTS_URL includes /v1/audio/speech, and set OPENAI_COMPATIBLE_TTS_PLAY=0 to isolate playback from synthesis.