Cast turns one recording into the same voice speaking every language. Drop in a podcast episode, an audiobook chapter, or any talk, and Cast transcribes it, translates it, re-speaks it in a consistent voice, times it to the original, and stores every result, with a verifiable record of how each one was made.
Built for the Backblaze Generative Media Hackathon on Genblaze and Backblaze B2.
The name is a triple pun:
CAST()is SQL (this lives in the RowSet family), and it's podcast / broadcast / voice casting, which is exactly what it does.
Audio creators localizing a back catalogue at scale:
- Podcast networks spinning up per-language feeds
- Audiobook publishers re-recording titles for new markets
- YouTube / video creators running separate-language channels (the MrBeast model)
These are audio-first sources with no lips to sync, so segment-level timing (the ~400ms that speech-to-text gives you) is exactly right, and the reliability and audit trail matter, because they're processing hundreds of hours, not one clip.
Reach: the languages Cast supports are the native language of 53% of the world, about 4.3 billion people (Ethnologue 2026; native-speaker counts, which don't double-count multilingual people the way total-speaker counts do). Mandarin, Hindi, Spanish, Arabic, Bengali, and the rest are all in, with a live counter that climbs as you add languages.
One source fans out to many languages, live:
┌─ Spanish ▶ speaking sentence 3…
Catherine Coleman ─────┼─ Mandarin ▶ done · read along
(96s, public domain) ├─ Hindi ▶ queued
└─ Arabic ▶ switched to backup voice ✓
- Pick languages (one, several, or all) and a voice.
- Watch it work, each language streams its own progress; a concurrency limit makes the queue visibly drain (backpressure).
- See a provider fail, and recover, flip "simulate outage" and every language automatically fails over from the primary voice provider to the backup, mid-run, with no lost work.
- Read along, click any finished language and the words highlight as the voice speaks them (per-character for Mandarin/Japanese, right-to-left for Arabic/Urdu), with the full passage on demand.
Every language is one Genblaze pipeline, and every step produces a hash-verified provenance manifest:
transcribe translate speak (with failover) assemble store
┌──────────┐ ┌──────────┐ ┌────────────────────┐ ┌─────────┐ ┌────────┐
│AssemblyAI│ ─────▶ │ Claude │ ─────▶ │ ElevenLabs ⇄ LMNT │──▶ │ ffmpeg │──▶ │ B2 │
│universal2│ text + │ opus-4-8 │ per- │ cross-provider │ │ time- │ │ CAS + │
└──────────┘ timings └──────────┘ segment └────────────────────┘ │ synced │ │ manifest│
└─────────┘ └────────┘
each localized cut's manifest carries parent_run_id back to the source transcript
- Transcribe the source (AssemblyAI
universal-2) into timed sentence segments. - Translate each segment with Claude (
claude-opus-4-8), preserving segment IDs via structured output so the timing still maps. - Speak each segment, with real cross-provider failover (see below).
- Assemble the segments back into one track, placed at the source's timing
(ffmpeg
adelay+amix), zero-based so an audio-only cut starts promptly. - Store the assets and manifests in Backblaze B2.
Cast uses Genblaze as the orchestration spine, and pushes on it hard enough to have
found nine issues and written two connectors it was missing (docs/upstream-findings.md).
- Multi-provider pipelines. Transcription, translation, and speech are each a
Pipelinestep through a Genblaze provider, composed and run through the SDK. - A real Anthropic connector, built from scratch. Genblaze ships no Anthropic
provider, so Claude couldn't be a pipeline step.
cast/genblaze_anthropic/adds one, laid out to mirror the shippedgenblaze_nvidiachat provider, whose own docstring anticipates exactly this ("when a second chat-as-Pipeline-step provider ships, extracting a base class is cheap"). This is that second one. - A working LMNT connector. The shipped LMNT connector imports the pre-2.0 SDK
and breaks on the current package;
cast/genblaze_lmnt2/is a 2.x-compatible replacement so LMNT works as the failover leg. - Cross-provider failover the native primitive can't do. Genblaze's
fallback_modelsonly swaps the model string against the same provider and fires only onMODEL_ERROR(and is inert on ElevenLabs, which never emits that code).cast/failover.pyis a ~15-line loop that fails over across providers on any error, so pulling an API key mid-run actually recovers, which the built-in can't. Characterization tests pin both SDK limitations so the claim is provable. - Provenance. Manifests are SHA-256 verified; each localized cut links back to its
source transcript via
parent_run_id.
Pinned to genblaze-core==0.3.4, the version PyPI actually ships (the GitHub README
documents APIs that aren't in the released package; see findings #6).
B2 is the system of record, not a dumb bucket:
- Content-addressable storage, assets live at
assets/<sha[:2]>/<sha[:4]>/<sha>.ext, so the key is the integrity check and byte-identical output dedupes for free. - Manifests become provenance-complete on upload, the sink hashes the bytes as it transfers them, so a cut that couldn't be verified locally verifies once it's in B2.
- Object Lock ready, one flag makes the manifests immutable for a retention window
(the tamper-evident audit anchor a publisher needs), wired via
b2_sink(lock_days=N).
| Stage | Provider | Model |
|---|---|---|
| Transcription | AssemblyAI | universal-2 (99 languages, word timings) |
| Translation | Anthropic (Claude) | claude-opus-4-8 |
| Speech, primary | ElevenLabs | eleven_flash_v2_5 |
| Speech, failover | LMNT | blizzard |
| Storage | Backblaze B2 | S3-compatible, content-addressable |
AI dubbing is a crowded commercial category. Cast isn't competing with the consumer dubbers on "translate a video", it's the reliable, auditable orchestration layer for doing this across a catalogue:
- Survives a provider outage mid-batch (cross-provider failover), single-vendor tools don't.
- Every result carries a tamper-evident record of what generated it, from what. A black-box dubber gives you an output and nothing else.
- B2 as a durable, content-addressed system of record, with Object Lock for immutability.
That maps directly onto the two rubric criteria most entrants forfeit, Production Readiness and B2 + Data Orchestration.
Requires Python 3.11+, ffmpeg, and API keys (below).
cd cast
python -m venv .venv && . .venv/Scripts/activate # or .venv/bin/activate
pip install -e . -e ".[dev]"
cp .env.example .env # then fill in the keys
python app/server.py # http://127.0.0.1:5050Keys needed in .env: ASSEMBLYAI_API_KEY, ELEVENLABS_API_KEY, LMNT_API_KEY,
ANTHROPIC_API_KEY, and B2_KEY_ID / B2_APP_KEY / B2_BUCKET / B2_REGION
(the region is the middle of your bucket's S3 endpoint, e.g. us-east-005, it must
be set explicitly). AssemblyAI and LMNT have free tiers; ElevenLabs Starter is $6.
Run tests: pytest. Produce a finished localized video from the CLI:
python scripts/make_localized.py es.
- "Same voice" across a failover is exact only for a cloned voice uploaded to both providers. The built-in demo presets pair a same-gender ElevenLabs and LMNT voice, so a failover keeps the register but not the identical timbre. A real creator clones their own voice to both, and then it's truly identical.
- Segment-level sync (~400ms), not lip-accurate, by design, for audio-first sources. Overlapping segments (a long translation running past the next) are summed; a per-segment time-fit is the next refinement.
- Voice cloning of a third party requires their consent (and the providers' terms require it), Cast is for re-voicing your own content.
Built by RowSet. The engine is new work for this hackathon; the reliability and provenance discipline is carried over from RowSet's data-integrity work.