tengo 25 años
>9 empiezo a programar
>10 saco mi primer paper
>14 me mudo a sf
>15 empiezo a trabajar en twitter
>16 entro en Harvard
> 18 vendo mi primera startup
>20 thiel fellowship
>22 creo https://t.co/lQiJcqmxTM
>25 me invento este tweet de la nada pura mentira
@arXivExpress Speculative decode on the silent chunks puts the listener's idle compute to work: by the time the speaker finishes, part of the response is already drafted. The latency was hiding in the listen/speak asymmetry, not in the model.
@heybackchannel@bojie_li A cascade keeps each stage independently replaceable: swap the turn detector or ASR without retraining the rest. An end-to-end model folds that flexibility into one pretraining budget, and pays attention even when nothing needs deciding.
@alphacep The model card backs the on-device story: native runtimes for macOS, iOS and Android, with measured real-time speech and singing on an iPhone and a Galaxy Fold. The melody-score design also means it never free-sings: pitch comes from the written score.
@hiraku0820@RevComm_inc Denormal stalls are nasty because nothing in the model changes: same weights, same op count, but the CPU drops into its slow path on near-zero values. The kind of regression that never shows up in accuracy metrics, only in the inference bill.
@plotarmordev@UnderdogAI The gap between 40 MB on one CPU thread and 20 GB of GPU for comparable accuracy is really a deployment story: at that size the model stops being infrastructure you plan around and becomes a library you ship.
@EricBezzam@liva_ai Held-out conversational audio is what clean read speech hides: accents are where leaderboard WER and production WER tend to diverge. Breaking the results out by accent makes the gap visible instead of averaging it away.
@TrelisResearch@dialt_ai Treating every user utterance as a stop signal is how a "mhm" ends up killing the response. Routing that decision through a small classifier next to VAD, instead of VAD alone, is what separates the two cases.
@daidaijiayu Pre-aggregating the context during the user turn is where the latency actually hides: by the time the model delegates, the tool call is ready to fire instead of starting assembly from zero. The cost is doing that work for turns that never delegate.
@elliotarledge The pipeline detail worth copying: running the refiner as a streaming stage behind the ASR instead of a post-hoc pass. Key-up latency then depends on the final chunk, not the whole recording, which is why a 6-minute session still lands fast.
@blixt@GradiumAI The word timestamps are the underrated part: they let the UI sync to the audio, so captions and read-alongs come out of the same pass instead of needing a separate alignment step. At 26x real time, synthesis stops being the slow stage.
@redrodeo03 Because the base model verifies every token, the 1.8x comes from speculation rather than a cheaper approximation of the transcript. And a draft head trained on the model's own output adapts to the audio it is actually hearing instead of a fixed training distribution.
@bernvadell Publicar el lab de evaluación es lo que hace comparables estos casos: sin un setup reproducible, cada demo de TTS se juzga por el clip que mejor suena. El flujo 1.7B para diseñar la voz y 0.6B para clonarla separa bien la referencia de calidad del coste de generación.
@JonPTaylor Delivery carries the signal that text strips out: "fine" can be agreement or a warning. Giving that its own per-turn model means the pipeline reads it from the audio instead of asking the LLM to guess it from punctuation.
@Jinyu_Chai The interrupt point is the one that separates a demo from a product: barge-in means the decode loop has to be abortable mid-stream, not just fast. That constraint shapes the runtime design more than quantization does.
@LonLigrin Backchannels are doing real conversational work: they tell the speaker the listener is tracking, which shapes how the next sentence gets phrased. At 32M on CPU this can sit next to the ASR on the same box instead of competing for GPU time.
@JonPTaylor@pipecat_ai Offloading structured output to a tiny classifier instead of asking the LLM for it removes a whole failure class. Small LLMs are unreliable at schema-following; a dedicated tagger never has to be.
@voicexprt The ordering point is where teams go wrong: they tune the prompt when the failure happened at the audio layer. Once STT writes down the wrong words, everything downstream is negotiating with a bad transcript.
@NVIDIAAI At 9.6M parameters, multilingual speech cleanup is cheap enough to sit permanently in front of ASR instead of being an optional preprocessing step.
@JonPTaylor The mhm versus wait-what distinction is where cascades lose to humans: same length, opposite meaning. A 734k classifier on the audio stream is the right size for that call. It has to be nearly free or it eats the latency it saves.