FreeAudioToText
Turn Any Audio & Video into Text in Seconds — 100% Free & No
-
1
nice work
-
1
Thank you @roni1233! Really appreciate the support! 🙏
-
-
1
I have learned an important lesson ! Like the free perspective
-
1
Thanks @startupsid! Glad our freemium and edge-architecture perspective resonated with you. Keep shipping! 🚀
-
-
1
13 mins for a 1-hour file on M2 Max is impressive, curious what stack or models you're running locally for that. I've been building something similar too, fully browser-based transcription with zero server processing. Same frustration with basic tools being paywalled tbh. Mine's here if you want to compare notes: kuberagent.com/tool/audio-to-text. Nice work either way ..
-
1
Thanks for asking, and congrats on building your client-side WASM tool as well! Browser-based processing is a fantastic approach for privacy.
Regarding our stack: The big speed gain without sacrificing accuracy comes from our hybrid model routing architecture running on an author-owned Apple Silicon M2 Max (32GB Unified Memory).
Instead of just relying on autoregressive Whisper for everything, we route Mandarin, Cantonese, and Asian/bilingual audio to Alibaba DAMO Academy's FunASR (Paraformer / SenseVoice) models. Because Paraformer is a non-autoregressive architecture, it infers orders of magnitude faster than Whisper while providing native timestamp alignment and punctuation restoration. For English and multilingual files, we use optimized Faster-Whisper paired with Silero VAD filtering to efficiently skip long silent pauses. Task orchestration is handled ephemerally via Cloudflare D1 and R2 at the edge!
-
-
1
“100% free is a strong hook.
Out of interest, have you looked at what would actually make users come back instead of just using it once and leaving? That repeat use case is usually where the real opportunity is.”-
1
Great question @Alfie! You hit the nail on the head regarding repeat usage.
Our retention loop relies on consistent utility without paywalls. Podcasters, journalists, and qualitative researchers transcribe files every single week. When they inevitably get hit by Otter ai or Descript's strict 3-file daily caps or 30-minute paywalls, they bookmark FreeAudioToText as their reliable, daily driver.
To capture long-term value from these returning users, we recently launched two retention levers: our Speech-to-Text API (allowing developers to integrate our fast, affordable engine into their Notion/Slack/Telegram bots) and embeddable iframe transcript cards for content creators. When the basic utility is frictionless, upgrade trust naturally follows!
-
1
“Really solid loop — ‘fallback when others cap out’ is smart positioning.
One thing we’ve seen: users often intend to come back, but don’t unless something actively triggers them.
Would be interesting to test where that intent drops off across segments — happy to share what that looks like if useful.”
-
1
That would be incredibly helpful, @Alfie! You're totally right that passive intent often needs an active trigger to turn into a habit.
I would love to see what those drop-off patterns and segment behaviors look like if you're open to sharing! Really appreciate you offering those insights, that kind of data is pure gold for an indie workflow. 🙏
-
1
Hi @double2 I built a buyer segment behavior report is there an email i can reach you on to send it?
-
1
Hi @Alfie, that is amazing! Thank you so much for taking the time to put that together. You can send it directly to me at support@freeaudiototext.com . I'm really looking forward to diving into the insights. Please let me know if there's anything I can help you with or any feedback I can provide in return! 🙏
-
-
-
-
-
-
1
Which four parameters? That matters more than the speedup number, because a couple of the usual faster-whisper knobs aren't free wins — beam_size=1 and int8 quantisation will together get you most of a 3000→783s move, and both cost you on accented speech, crosstalk and noisy rooms. Which is exactly what interview and meeting audio is.
Worth running WER on a held-out set at the old settings vs the new ones before calling it a pure optimisation. VAD filtering is usually the one that's close to free.
Also, is 783s a warm-worker number? If the model unloads between jobs, the first job after a quiet period pays the load cost — and on a free tool with bursty traffic, that's the number most of your users actually feel.
-
1
Spot on! You are asking the exact right engineering questions @jorvexis.
On the 783s speedup & WER: We absolutely avoided the "cheap speed wins" like reckless int8 quantization or forcing beam_size=1 on Whisper, which destroy accuracy on accented and noisy speech (like Zoom interviews). The speedup from 3000s -> 783s was achieved by implementing a hybrid routing pipeline: replacing Whisper for Chinese/Asian audio with non-autoregressive FunASR (Paraformer-zh), which is natively faster without WER degradation, and introducing aggressive Silero VAD filtering before passing chunks to Faster-Whisper for English. This skips dead air and room noise without chopping speech boundaries.
2. Warm Worker vs Cold Start: Yes, 783s is a warm-worker number! Because we run our Python inference worker on dedicated, author-owned Apple Silicon hardware (32GB RAM) continuously polling Cloudflare D1, the model checkpoints remain loaded in unified memory. We don't spin down or unload models between jobs, so our users never pay a cold-start model loading penalty even during bursty traffic!
-
1
Thanks for the detailed reply, that answers both. Routing to a different engine rather than reaching for the quantisation knobs is a good call.
-
-
1
Making the core transcription free is a strong acquisition hook, but the real challenge is whether the paid analysis is valuable enough to cover the compute cost created by unlimited free usage.
I’d be interested in two numbers: the average cost of processing one hour of audio, and what percentage of users upgrade to the $1.99 report. Without those, “free and unlimited” could become a growth feature that quietly destroys the economics.
Also, auto-deleting files within 24 hours is good, but I’d make the privacy claim more concrete by explaining whether audio is used for model training, whether transcripts are stored separately, and where processing happens. “Privacy-first” is easy to say; specifics are what build trust.
-
1
Thanks for the thoughtful feedback @SongTrailer! These are the exact economics and ethical principles we care about most.
The Economics (Zero Marginal Cost): Why doesn't unlimited free usage destroy our unit economics? Because our inference engine runs on a dedicated, author-owned Apple Silicon M2 Max Mac that we already own for our daily engineering work. Our incremental cloud compute cost is literally $0 (Cloudflare Pages, D1, and R2 free tiers easily cover millions of database reads and ephemeral file passing). Therefore, even a single $1.99 Waffo report upgrade covers our bandwidth/storage overflow for hundreds of free users!
Concrete Privacy Claims: You are 100% right that "privacy-first" needs specifics. To be completely transparent: • No AI Training: Uploaded audio and generated transcripts are never used to train, fine-tune, or prompt any AI models. • Isolated Processing: Audio is processed locally on our secure, private hardware worker—never sent to third-party transcription API farms. • True Ephemeral Purge: An automated Cloudflare edge cron job permanently and irrevocably deletes all R2 object storage files and D1 database records after exactly 24 hours.
-
-
1
The free transcription angle is an interesting distribution choice.
What would convince you that the paid analysis layer is valuable enough to support the business, rather than users only coming for the free transcription and leaving?
-
1
Thanks @Aryan! Our distribution strategy is focused on building top-of-funnel SEO dominance (ranking on Google for organic terms like "free audio to text").
Because our infrastructure architecture (author-owned Apple Silicon + Cloudflare Edge) brings our marginal compute cost down to near-zero, we don't need a traditional 5% SaaS subscription conversion rate to survive. Even a 0.5% to 1% conversion rate to our $1.99 AI Structural Analysis Reports and developer API usage creates a highly profitable, self-sustaining indie business model. Offering genuine, unlimited free value is our best marketing engine!
-
1
Appreciate the context.
It’ll be interesting to see whether the free workflow naturally creates enough demand for the paid analysis layer over time.
-
-
-
1
This comment was deleted a month ago
Comment