TTS

Narrating a HyperFrames docs video

Anything that will sit on hyperframes.heygen.com uses one narrator, so the site does not sound like two products.

This is the voice every user-journey film on the docs site already uses. Falling back to local Kokoro because a key was not to hand produces a film that sounds wrong beside the others — three docs videos were built that way and had to be re-voiced. If you cannot reach ElevenLabs, say so and stop rather than substituting a different voice.

Use another voice only for a documented reason, and write the reason down.

Available routes

Gemini is an explicit alternative to the automatic provider order below. A request to use Gemini already chooses the provider; do not redirect that user to HeyGen sign-in. Read the Gemini section for its credential requirement.

Self-contained HeyGen (no CLI) — scripts/heygen-tts.mjs

The published hyperframes tts CLI synthesizes locally with Kokoro only. When you want HeyGen specifically — best quality plus word timestamps in one call — use the skill’s bundled script, which calls the HeyGen v3 REST API directly and needs no CLI provider plumbing:

The script resolves a HeyGen credential the same way the CLI does — first source wins: $HEYGEN_API_KEY → $HYPERFRAMES_API_KEY → a project .env (auto-loaded, walks up ≤5 dirs) → ~/.heygen/credentials (shared with heygen-cli; $HEYGEN_CONFIG_DIR overrides the dir). An OAuth login is sent as Authorization: Bearer; an API key as X-Api-Key; both include X-HeyGen-Source: cli. OAuth CLI users can consume the web-plan free allowance (10 min/month) before paid usage; API keys follow normal API billing. If the only credential is an expired OAuth token it stops with a hint to run npx hyperframes auth refresh.

• Voice: --voice <id> must be a starfish voice_id (--list, or GET /v3/voices?engine=starfish). v2-catalog ids are rejected with HTTP 400. Omit --voice (English) and it defaults to Marcia (05f19352e8f74b0392a8f411eba40de1, a fixed default so the choice is deterministic). Non-English with no --voice falls back to the first matching catalog voice.
• Output: .wav → transcoded to 44.1k mono via ffmpeg; .mp3 → raw bytes (no ffmpeg needed).
• Words: --words <path> writes the flat [{id,text,start,end}] shape below, drop-in for the captions pipeline. HeyGen’s <start>/<end> boundary sentinels are filtered out and ids are re-contiguous.
• Non-English: --lang <code> (anything but en) is sent as the request language.

When to use which provider

Gemini narration

Use the shared audio engine, not hyperframes tts. Authenticate with either:

• GEMINI_API_KEY or GOOGLE_API_KEY (first set key wins).
• A service-account JSON file at GOOGLE_APPLICATION_CREDENTIALS, or injected JSON in GCS_CREDS. Install google-auth requests in the Python 3 environment used by the helper. A configured file takes precedence over injected JSON.

API keys take precedence over service accounts: unset both key variables to use OAuth. Never put credentials in a request file or composition. The helper obtains a fresh OAuth token for each generation with the generative-language.retriever scope. The quota project resolves from GOOGLE_CLOUD_PROJECT, then GCLOUD_PROJECT_ID, then the service-account JSON. User ADC files and metadata-server authentication are not supported.

Both routes call the Gemini Developer Interactions API, not Cloud TTS or Vertex AI. Those APIs have separate model catalogs and access requirements.

Save this as audio_request.json in the project:

The engine saves assets/voice/intro.wav, measures its duration, and transcribes it into voices[].words in audio_meta.json. Check that every requested line has audio and nonempty word timings before building a captioned video. Review the timings against the actual audio; transcription is estimated alignment, not native TTS timestamps. Do not distribute words evenly across a clip.

• Models: gemini-3.8-flash-tts (default), gemini-3.8-flash-lite-tts, gemini-3.1-flash-tts-preview, gemini-2.5-pro-preview-tts, and gemini-2.5-flash-preview-tts. Use these exact Developer API IDs; Cloud TTS aliases such as gemini-2.5-flash-tts are not accepted.
• Voice: Kore by default; pass another prebuilt voice or an existing custom voice ID for 3.8. Older models require prebuilt voices. Creating or replicating voices is outside this helper.
• Delivery: Put directions in style, not in spoken text. Each line can override style. 3.8 uses structured annotations; older models receive a delivery prompt before the transcript. Check that directions were not spoken. Use style for pacing; numeric speed must be omitted or 1.
• Audio: 3.8 returns a complete WAV. Older models return mono 16-bit PCM, which the helper wraps as WAV at the returned sample rate without resampling. No generation calls run during playback or rendering.
• Timing: This adapter requests no native word timestamps. It uses the existing transcription pass. lang selects transcription language; Gemini infers speech language from the text.
• Workflow adapters: Product-launch, faceless-explainer, and PR-video audio scripts accept --provider gemini --voice Kore --tts-model gemini-3.8-flash-tts --style "Warm and clear".

An API error is reported as a failed line; an explicitly chosen Gemini voice never silently falls back to another provider. Check the engine’s anomalies and output metadata, not only its exit code.

API contract: Google’s speech generation guide (opens in a new tab).

ffmpeg requirement

HeyGen + ElevenLabs return mp3. The bundled HeyGen helper transcodes to wav when --output ends in .wav (the default and what downstream ffprobe + Whisper expect). If you’d rather skip the transcode, pass -o file.mp3. Without ffmpeg on PATH, wav output from cloud providers fails; the local Kokoro CLI writes wav directly.

Voice selection (Kokoro)

Default af_heart. Curated picks:

Run npx hyperframes tts --list for the bundled set.

Multilingual (Kokoro voice prefix → language)

The first letter of a Kokoro voice ID picks the phonemizer language; --lang overrides auto-detection.

Valid --lang codes (only needed to override the voice’s auto-detected language): en-us, en-gb, es, fr-fr, hi, it, pt-br, ja, zh.

Non-English phonemization requires espeak-ng system-wide (brew install espeak-ng / apt-get install espeak-ng).

Speed

• 0.7-0.8 — tutorial, complex content, accessibility
• 1.0 — natural pace (default)
• 1.1-1.2 — intros, transitions, upbeat content
• 1.5+ — rarely appropriate, test carefully

The hyperframes tts command honors --speed for Kokoro. Provider-specific helpers document their own pacing controls.

Long scripts

Past a few paragraphs, write the text to a .txt file and pass the path. Inputs over ~5 minutes of speech may benefit from splitting into segments.

HeyGen word-timestamp shape

When --words <path> is passed to a HeyGen call, the file is written in the same flat shape transcribe produces — drop-in compatible with the captions pipeline:

For ElevenLabs / Kokoro, run npx hyperframes transcribe narration.wav --model small.en to get the same shape.