Wondercraft: turning speech generation into a production workflow

The blog already exists. The recording does not.

A company has already paid a content team to produce a good blog post. Reaching listeners means taking on another production process. Youssef Rizk’s early example for Wondercraft, an AI podcast builder, starts with that existing post and converts it into an audio-friendly format. The interview itself illustrates the overhead: the participants reportedly spent about 50 minutes setting up the room, on top of coordinating schedules and preparing the conversation.

In this September 2023 conversation, Rizk describes a deliberate application-layer business: assembling AI capabilities into a product for a particular task. His background in forward deployment engineering combined software development, solutions architecture, contracts, sales, and customer interaction. He endorses that breadth as preparation for starting a company.

The broader ambition is to expand access to existing content. Audio lets someone listen while doing household tasks. Rizk sketches blog-to-video, podcast-to-blog, and podcast-to-Twitter-thread conversion as possibilities beyond the initial builder. Personalized meditation is another proposed use case; Calm and Headspace illustrate the category, without being identified as customers.

From source text to a publishable episode

The core pipeline has two stages: generate or supply a script, then synthesize its speech. At the time, Wondercraft uses ElevenLabs for text-to-speech, the conversion of written text into spoken audio. Creators can supply finished scripts. The surrounding tools aim to keep podcast production from becoming a recurring distraction.

Those tools cover several distinct responsibilities: - Correct pronunciation: specify how a word should sound when the generated voice gets it wrong. - Reuse show settings: organize episodes under podcasts, with defaults for intros, outros, music, and speakers. - Publish and distribute: generate show notes and host the podcast’s RSS feed—the episode feed creators can submit to Spotify. - Adapt to other formats: produce in 28 languages and create social video clips. Clips remain in beta, while interview-style templates are under development.

Newsletter conversion makes the automation concrete. The newly announced Wondercraft Zapier integration detects a newsletter appearing in a feed, retrieves it, and generates an episode. It can publish automatically, but generation and publication can remain separate: a creator can edit the result before releasing it or schedule a later release.

Review before publication does not resolve every update problem. Automatic propagation of later newsletter edits into an already published episode remains unspecified. The manual newsletter conversion is estimated at about two minutes, with no newsletter length or other measurement conditions supplied.

Curation determines whether audio is worth the time

Publishing automation leads to a discussion of cadence. The interviewer recommends publishing at least weekly; the product-side argument emphasizes a dependable schedule suited to the audience. Hacker News Recap reportedly reached Spotify’s top 30 technology podcasts, which its producer partly credits to daily publishing. The interviewer points to its Hacker News launch. The relative contributions remain unresolved, and proposed ranking mechanisms involving download volume or listening time are speculative.

The team had dropped Product Hunt Daily and identified Hacker News Recap and PG Essays as its two most popular in-house shows. Proposed audio feeds for subreddits such as Life Pro Tips and WallStreetBets expose a different constraint: some source material depends on images. Memes do not fit straightforward narration; improved video support is floated as a future option.

The interviewer’s personal Discord digest exposes the cost of listening to noise. He scrapes several communities into a daily newsletter but hesitates to convert it into audio: he can scan text and skip irrelevant discussion faster. Hacker News already supplies community curation, and Reddit’s hot posts offer a related filter, though he considers Reddit noisier. Upstream selection reduces the listening burden before speech generation begins.

PG Essays shows why long-form audio also needs navigation. It uses different voices for quotations, and its adaptation of How to Do Great Work runs roughly an hour. One participant listened instead of reading; the other says the podcast led him to consume more essays. The interviewer nevertheless wants an opening overview and explicit part-by-part signposting. The producer says chapters exist and introductions could improve, while maintaining the essays largely as written. The editorial challenge is to preserve the writing while helping listeners keep their place.

Academic papers pose a similar adaptation problem. Using Papers Read on AI with Rob as an example, the interviewer requests selective handling of citations, headers, and tables. Document elements that help a reader can become distracting narration. The useful transformation happens before synthesis: deciding what should be spoken and how.

Dubbing preserves who speaks when

The next announcement extends conversion across languages. AI dubbing accepts a podcast episode and generates translated speech using the original speakers’ voices. If video accompanies the recording, Wondercraft fits the translated audio to the original speaking chunks.

Speaker-turn alignment means keeping each translated turn within its corresponding part of the conversation. One participant’s dubbed voice should not continue after the video moves to the other participant’s turn. The respondent identifies timing alignment as the difficult part. Speeding up a translation that runs long is the interviewer’s suggestion; the response acknowledges adjustments without disclosing the method.

Lip synchronization—matching mouth movements to the translated speech—is a separate, unimplemented feature at the time. The product-side judgment is that listeners seeking a podcast in another language care primarily about access to its content. Matching conversational turns is already part of the workflow; matching lips remains a possibility.

Test willingness to pay before polishing the platform

The origin story returns to Moonshot, the team’s earlier venture enabling investment in people. Its work was primarily legal and regulatory. The team concluded that people did not want the product and that the work did not suit its strengths. It ranked possible pivots by conviction and chose content conversion, starting with a podcast builder.

The first prototype was a rough, single-page React form without authentication; anyone who found it could theoretically generate unlimited audio. Its purpose was to expose the idea to rejection quickly. The later test-with-example flow pursued the same immediacy by minimizing clicks between login and generated audio. Even two or three clicks felt excessive, prompting a proposal to put the experience on the landing page.

A Twitter post about the Hacker News daily project brought inbound interest and triggered a payment test. To ship within an hour, the team put Stripe in one app and redirected buyers to a separately hosted generation app that still lacked authentication. The prototype reportedly ran on one EC2 instance. With an arbitrarily chosen price of $50 per month, the team reports making about $3,000 in one day. The immediate test was whether anyone would pay before further investment.

The moat question starts with the customer

What protects an application built around someone else’s speech API? The response distinguishes a general capability from a complete customer workflow. A speech API can support many tasks without supplying the editor, publishing controls, music, and reusable settings a podcaster needs. Owning the application also lets Wondercraft choose the underlying speech engine.

The proposed defense is to build the right product for a specific user. Defining the company primarily through AI risks organizing it around implementation rather than customer needs. Product quality, network effects, distribution, execution speed, and early arrival are offered as possible advantages, not demonstrated protections for Wondercraft. The interviewer adds that founders need an authentic answer for themselves: they are committing their time, beyond investors’ money.

The respondent turns the question toward API providers, too. Paying for training, hiring researchers, and launching early can create an initial advantage, but visible demand attracts competitors—assuming they can recruit the necessary talent. His position is that a sufficiently large market can support several good products, with quality judged against the needs of each product’s customers.

Pleasant speech and disclosed synthesis

The preferred voice changes with the listening task: - Meditation: a slow, deep, relaxing delivery may suit the content and listener. - News narration: Hacker News Recap uses Anna, the default voice described as professional, neutral, and news-anchor-like. - Conversation: participants favor pauses, incomplete utterances, laughter, and overlapping speakers as signs of a natural exchange.

Wondercraft describes narration as its strength, with multi-host tooling still under development. The interviewer questions how much human resemblance an information service needs. In a played Procreate Dreams excerpt, Anna sounds excessively breathy to him. He tolerates that mistake because the show discloses its AI origin and he listens for information. The respondent defends natural delivery as a way to make listening pleasant, while recommending disclosure and reporting that the podcast’s Spotify page identifies it as AI-generated.

Wondercraft’s stated workflow transforms information supplied by the user: a blog post, an actual podcast, or another artifact. Watermarking output remains an intention awaiting a standard, rather than an implemented feature; the respondent also acknowledges not being fully current on that work.

Choose speech tools for control as well as sound

The tooling discussion distinguishes several kinds of value: - Transcript-based editing: Descript earns praise for editing video through its transcript and for Overdub. Wondercraft’s respondent emphasizes automation for retrieving source material and forming scripts, beyond editing alone. - Hosted speech generation: the options discussed include Amazon Polly, Google Text-to-Speech, Microsoft Cognitive Services, ElevenLabs, and Play.ht. - Open-source speech systems: the interviewer names Tortoise-tts, Pyttsx, Coqui, and Larynx, without a comparative evaluation.

The respondent prefers ElevenLabs’ voice quality but describes Play.ht as a more comprehensive platform, including pronunciation controls. That listening preference has a material version caveat: he is unsure whether Play.ht’s newer model has shipped, while the interviewer says it has. The comparison does not identify which model he evaluated.

A synthetic Joe Rogan–Steve Jobs conversation raises a separate question: cheap production does not establish audience value. The respondent regards the conversation as a viral gimmick he would not personally keep listening to, and questions the value of easier generation in an already crowded content market. Celebrity voices remain a possible future addition for Wondercraft, with licensing and payment considerations, rather than a priority.

SoundStorm’s generated dialogue impresses the respondent through its humanlike utterances and laughter. He says it has not been open-sourced at the time. Asked about reimplementing it, he favors delaying investment in compute until the product has established market demand. The qualification is consequential: compute spending should justify the runway it consumes. A company building a new speech engine might reasonably invest earlier.

After the demonstrations, the evaluation criteria become more concrete. Naturalness and willingness to listen matter, but production requires predictable delivery: - Duration control: a promised 30-second segment needs to fit that duration within an agreed tolerance. - Tone control: a requested emotion should produce reliable delivery, rather than an awkward approximation. - Customization: creators need usable control over how the output sounds. These are desired capabilities. The discussion establishes neither a precise duration tolerance nor a measured reliability result.

Dubbing adds an evaluation problem creators may be unable to solve themselves: plausible foreign-language audio can sound wrong to native listeners. Regional accents complicate judgment further. The product-side response says professional translators already participate in quality assurance for dubbing, with that capability expanding. A Latent Space participant offers Chinese listeners who translate episodes as potential reviewers; that remains a proposal.

The customer interview: save time and preserve editorial control

The closing conversation reverses the interview and examines Latent Space’s production needs. Its podcaster describes three formats with different workflows: - Founder interviews capture a person and company at a particular moment. - Breaking-news Twitter Spaces aim for an in-depth day-one explanation and require substantial editing. - Fundamentals episodes are planned nearly minute by minute and intended to remain useful years later.

Twitter Spaces show why an expensive production step can still be worthwhile. The podcaster reports cutting roughly two hours to one, using Descript to remove silences and filler words. That editing effort buys spontaneous participation without booking: Simon Willison is cited as having appeared three times without scheduling. For show notes, Small Podcaster is a roughly 100-line Python script that sends transcripts to Anthropic.

The ability to build tools does not eliminate willingness to pay for saved time. The podcaster wants RSS-to-podcast conversion and a less laborious clip workflow. He has not yet tried Wondercraft’s RSS automation, and his adoption condition is clear: when the initial result is unsatisfactory, the creator needs enough control to fix it.

Marketing is harder for him to delegate. Recognizable guests such as George Hotz and Chris Lattner can attract interest through their reputations; unfamiliar guests need an explanation of why listeners should care. Making that case requires his own writing voice. AI-generated tweets defaulting to emojis and hashtags illustrate the mismatch between producing promotional text and expressing his judgment.

The final invitation is to start a podcast or expand an existing one’s accessibility through dubbing. The advice to engineers and founders echoes the early payment test: develop your own conviction, launch, and keep spending restrained. Others’ claims about who has a moat should not substitute for deciding whether the product is valuable.

The recording’s outro also points listeners to an additional 30 minutes of YouTube video covering more Wondercraft features, use cases, and its origin story.