Why TTS Models Now Look Like LLMs — Samuel Humeau, Mistral

Talk transcript

1. 0:00

   [on hold music] So I'm, I'm from Mistral AI.

2. 0:17

   Uh, and we are going to talk about, uh, speech generation and text-to-speech. Uh, there is an occasion, uh, we released last week our first text-to-speech model and it's open source, so I really encourage you to, uh, check it out.

3. 0:29

   It's an extremely strong text-to-speech model. We are very proud of it. Um, and for this occasion, uh, I thought we could review some of the recent trend in text-to-speech architecture since there is a, a dominant, uh, trend, uh, emerging these days, although this can change like, uh, ve- very quickly.

4. 0:52

   Um, and, uh, so this talk is slightly academic and addressed to people who wants to know a bit more about how you, uh, do text-to-speech. Uh, this being said, we have a few years before the machines do all the science for us, so we might enjoy it, uh, today.

5. 1:09

   Um, I'm Sam. Uh, yeah, I work at Mistral as AI scientist. Before I was at Facebook FAIR when it was called, uh, Facebook.

6. 1:19

   Uh, and Mistral, a few words about the company. It-- We are a frontier lab. We have-- we've been founded, uh, a couple of years ago. Uh, we produce frontier model, uh, but we're also a B2B business.

7. 1:32

   We, we help organization, uh, in their AI transformation, which is kind of a buzzword, but literally every company is transforming with AI. We help them by providing them tools, product, and, uh, dedicated people to help them in their custom needs.

8. 1:48

   Um, back to the text-to-speech. So there, there are a few offline use case of, uh, speech generation, like the, the famous listen to the blog or listen to the article.

9. 1:59

   But nowadays, the, the king use case for text-to-speech is, uh, its usage within agents. And in particular, it's used to interface, uh, with a chat agent, typically in a pipe like this, uh, where you have a central, uh, chat agent that does text-to-text but does it extremely well and you want to talk to it, so you add

10. 2:20

    a, a speech-to-text, and you want it to speak to you, so you add a text-to-speech.

11. 2:26

    Um, as everybody in this conf will tell you, the latency is key here. Uh, so you can reduce the latency on the left by having the speech-to-text done in real time so that when you detect the end of turn, uh, you already have the transcript, it's already done.

12. 2:41

    Uh, and we are gonna focus a bit on the right side today. Uh, it's also very important that as soon as you have the first audio packets, you, you, you start to, um, to voice them out.

13. 2:54

    This way the perceived latency is lower. In fact, since your LLM can stream, uh, some text to you, uh, actually what you ultimately want is something like this if, if you're gonna interface a chat assistant, uh, which is a, a real time text input, text-to-speech where, uh, as soon as you have the first token of the LLM,

14. 3:16

    uh, the, the machine starts to speak. We're going to talk, uh, about it in the, in the end of the talk. I want to focus a bit, uh, at the beginning at the output side, uh, and what it is-- what it means to stream audio.

15. 3:30

    So to illustrate this, I have, uh, this app that I vibe coded for the occasion. Um, and so we are, we are going to, uh, to use this text-to-speech model that we released, that I mentioned, and we are gonna hear Paul.

16. 3:46

    So Paul is an actual human being that sounds like this.

17. 3:50

    The persistent anxiety that fills the rest of my life is calmed for as long as I have the flavor of something good in my mouth.

18. 3:58

    Um, so this is, uh, uh, like some actual recording on some, uh, actual person named Paul, and we are copying his voice. Uh-

19. 4:08

    And so with the sunshine and the great bursts of leaves growing on the trees, just as things grow in fast movies, I had that familiar conviction that life was beginning over again with the summer.

20. 4:18

    Um, so let's focus first on what's happening here. Uh, as you can see-

21. 4:23

    And so with the sunshine-

22. 4:24

    We are copying the voice and the, the first audio packet happens first, and we can start to emit audio, which greatly reduce the perceived latency even though the full computation of the, the audio happens like a few seconds later.

23. 4:39

    And so with the sunshine and the great bursts of-

24. 4:40

    So in... If you use it in an agent, so here I, I crafted, uh, a small agent using a, a speech-to-text, uh, one of our, uh, LLM and, uh, this very text-to-speech.

25. 4:52

    So we, we can speak to Paul, um, and... Hey, Paul, uh, can you tell me what's the title of the session at twelve twenty, please?

26. 5:04

    The session at twelve twenty PM is titled Ricci Mini: Giving a Body to AI by Andres Marafioti.

27. 5:11

    And what was the session at eleven fifteen, please?

28. 5:19

    The session at eleven fifteen AM is Beyond Transcription: Building Voice AI That Actually Understands Conversations by Hervé Bredin.

29. 5:29

    Did you enjoy it as much as I did? [laughs]

30. 5:33

    I don't have personal experiences or emotions, but I'm glad you enjoyed it. [laughs]

31. 5:38

    That's all I can do. Um, so the important thing here is that since the audio packet arrived first, you still have a, a decent latency, and you can enjoy the conversation with the agent, uh, despite the fact that the audio is still, uh, not generated fully.

32. 5:56

    And so we're gonna dig... Oh, yeah, sorry. Uh, I want to make, uh, one digression. So I, I mentioned the, the voice cloning here. Uh, this model can Like, only need a few second to clone the voice of, uh, someone.

33. 6:09

    Uh, so again, this is how Paul sounded and this is how we generate text

34. 6:15

    All day long I could And so with the sunshine and the great bursts of leaves growing on the trees

35. 6:17

    It, it really sounds like ... It's also very good at inferring how a person would speak in an, in, uh, another language. So for example, this is a voice, uh, a [REDACTED:origin] voice.

36. 6:30

    Maybe.

37. 6:31

    Ryan Coogler, réalisateur et scénariste des deux Black Panth-

38. 6:35

    And, like, if we generate

39. 6:36

    And so is the sunshine and the great burst of leaves growing on the trees, just as things grow in fast movie.

40. 6:43

    Yeah, so we can clearly recognize her, and we can clearly recognize the very strong [REDACTED:origin] accent, which as a [REDACTED:origin] myself, I do enjoy. Um, I can, I can even clone my own voice.

41. 6:54

    So this is how I sounded during the recording.

42. 6:58

    Hi, this is Sam. And, uh, this is how I sound. And so with the sunshine and the great bursts of leaves growing on the trees. It works pretty well.

43. 7:08

    So this way in my time of dis- delusion and at the peak of my ego, I can discuss with myself, uh, on, uh, on a complicated problem, which is nice.

44. 7:19

    Um, and so it, it's becoming so easy to, uh, impersonate a voice that it's, uh, it's becoming very easy to, to configure. Um, so it's a small digression, but currently actually a lot of large company, they do have a concept of vocal identity, and they do care in their branding about how they sound, uh, in their advertisement

45. 7:39

    in particular. But I think this, uh, concept will be becoming more mainstream and just as like a lot of company, uh, define how their website appear as their brand identity, uh, it would be the same for the voice, uh, identity.

46. 7:55

    Um, oops, sorry. Uh, back to how we do it, uh, in general. So, uh, right, this is, this is, uh, an audio. Uh, physically it's the pressure of the microphone that we measure, uh, from time to time, like, uh, several thousand of times per second.

47. 8:12

    So it looks like this. And historically, to generate the audio, uh, there have been a lot of, uh, attempts, a lot of systems like i- in the prehistoric time you have like stitching of, uh, words that were spoken like, uh, in the [REDACTED:origin] train, uh, system SNCF for, for those who knows.

48. 8:30

    Um, and then a neural generation arrived, uh, at some point the trend was to generate each sample one after the other. Uh, then, uh, another era was generating the whole audio at once.

49. 8:41

    But as we can ... Uh, as we saw, it, it's very interesting to have the beginning of the audio generated first so that we can start to play it out.

50. 8:49

    So it seems that most labs have converged to some common patterns, and obviously the, the first one is inspired by large language model. We are trying to, uh, transform the problem as a language modeling problem because humanity is extremely good at modeling sequences of token.

51. 9:07

    So pretty much, uh, everybody is using an autoregressive decoder backbone and, uh, generate audio one piece after the other. Now, um, as I said, we really don't want to generate one sample after the other.

52. 9:24

    So what we want to do is generate like patches of audio one after the other.

53. 9:30

    Um, so the expected system looks like this where you, you have an encoder that transform like a, a frame of audio, something like 80 millisecond into something that ideally is a token because humanity is very good at, uh, modeling sequences of token, and then you have a decoder that does the opposite.

54. 9:49

    Uh, now for text it's pretty easy because transforming the text into tokens, well, it, it's easy, right? You can take like words, uh, as token and it, it works pretty well even though we, we do much better.

55. 10:00

    Um, for audio it's much harder because one token doesn't have a lot of in- information. Like one token of a vocabulary of a thousand is 10 bits of information and the audio requires much, much more, uh, like a much larger bitrates.

56. 10:15

    For example, a standard quality MP3, that's 200 kilobits per second.

57. 10:20

    And so in order to transform this into a sequence of token, like, and not have thousands and thousands of tokens, uh, we need to somehow compress it and reduce the size of it, maybe drop what's not needed.

58. 10:35

    Uh, an interesting point of comparison is tec- text captioning, uh, because if you drop all the acoustic information and you, you just focus on the text like, uh, with a subtitle track, uh, you actually drop most of the information.

59. 10:48

    It's a massive reduction and, and you only have, uh, a few bits per second, uh, remaining. So here in this, in, in this demo I use, uh, our real time speech-to-text to measure my bitrate in terms of, uh, tokens per second of text.

60. 11:04

    And, um, I, I'm a very competitive person and a very ... I'm very good at speaking. Uh, yet, uh, I'm barely 15 bits per second of, uh, actual information.

61. 11:14

    You, you can try to, to, to, to beat that but like in the grand scheme of things, compared to 200,000, uh, bits of informations per second, that's not a lot.

62. 11:25

    Uh, obviously, uh, we want to use something that allow us to recover like acoustic, uh, features like the voice and, uh, and other aspects and not just the, the semantic information as in the text.

63. 11:39

    So the codec that are used, uh, typically reduce the audio to about a few thousand, uh, bits, uh, per second.

64. 11:49

    Um, in our case for instance, uh, we treat the problem with ... Uh, sorry, we cut the audio as, uh, with pieces of 80 milliseconds, so 12 frame per second, and we transform each frame into several tokens, like 37 in our case.

65. 12:05

    So we reduce the problem to about 500 tokens per second. I'm not, uh, gonna dig too much, uh, on how we do train these codecs, but obviously first we train them.

66. 12:17

    Uh, we train them by reconstructing a very large set of audio and using a bottleneck here. Uh, the, the training procedure constrained, uh, the reconstruction to go through a step where each frame is, uh, decomposed into several tokens.

67. 12:38

    Uh, to do this, typically it's guided so that the model drops the information that is useless and only retain the one that is, uh, useful. And so we guide it via some losses, uh, reconstruction losses, uh, adversarial losses and, uh,

68. 12:58

    particularly like for, for the, for some of the tokens, we try to make sure that they contains the text information so that you can reconstruct the text from it.

69. 13:10

    Still 500 tokens per second is a lot of tokens. Uh, and you, you could put them one by one aligned as a sequence like this, uh, but it would make a lot of step of the main transformer that is, uh, at the core of the system, which is usu-

70. 13:28

    It's huge. Um, in our case, it's 4 billion parameter, [clears throat] sorry, which is, which is still a lot, even though it's not like extremely big now. Um, what most people do then, uh, is,

71. 13:45

    uh, have, having one step of the backbone per frame and a smaller model here, uh, typically a dev transformer that, um, recomputes all the, the tokens of one frame, uh,

72. 14:00

    at, at each step. So this way you still have a lot of tokens, you still carry a lot of information, but the computation is much faster.

73. 14:09

    So this is the, the main, uh, pattern that we see. Uh, even though for that last bit, like the model we released does not follow this pattern, we actually, uh, defer on that part.

74. 14:22

    I'm not gonna dig too much on it, but just so you know, each frame, which is represented by 37 tokens in our case, we do generate these 37 tokens at once using a diffusion model.

75. 14:33

    So it's slightly different from the, the vanilla, uh, text-to-speech nowadays. I encourage you, by the way, to read our technical report, which contains all this information.

76. 14:44

    Uh, also it's a pretty cool, uh, use case of, uh, flow matching, uh, models which is similar to diffusion model.

77. 14:53

    Uh, now the, the main path is conditioning actually, uh, because so far we are just generating audio, but we are not conditioning it on text, so it's not really a text-to-speech, it's just a speech.

78. 15:04

    Um, and to conditioning the... For conditioning, there is way more variance across labs and papers and implementations. Um, you have typically two categories. Uh, there are the, the people who focus on, uh, producing the audio once you have the, the text, and some of them who focus on having a stream of text.

79. 15:30

    Typically, the first category will tend to provide all the context at the beginning and then produce the audio as we saw. Um, and typically, the second category will also add some context as the audio is, is produced.

80. 15:46

    Uh, the model we release is in the first category. So what we do is we provide the audio of the voice we want to clone, so a few seconds, then the text to pronounce, uh, and that's our context in our case.

81. 16:03

    Um, yeah, regarding the latency, so it, it, it's, it's pretty fast, uh, if you remove the network and with a single GPU you have 17 milliseconds between the moment where you input your text and the moment where you have the first audio you can play.

82. 16:22

    Uh, regarding real-time text input stream, uh, which, uh, is our next step, uh, for us, uh, there is no, uh, real... There, there is no clear, there is not a clear winner.

83. 16:35

    First, uh, it's still possible, you know, like to, to generate independently the text and, and stitch them out, but obviously you will have, uh, a lot of continuity problem.

84. 16:45

    Um, and there are several patterns. The, the two main ones are people who interleave audio and text, so as soon as there is a new text, they, they put the text in the same, uh, layer.

85. 16:58

    Uh, and some other who have a dual stream architecture where you have a stream of audio and a stream of text, uh, and you kind of blend them together during the, um, the inference.

86. 17:10

    Um, how am I doing on time? That's... There's two minutes remaining. Thank you. Uh,

87. 17:17

    the takeaway is, uh, check our, uh, open source model please. Uh, read the, the technical paper and I hope you learned a few things today. [audience applauding]

88. 17:31

    We do have two minutes. Uh, how you spend them is up to you. Yeah.

89. 17:35

    Um, you said that your model first takes all of the text and then produce the audio. But on the example that you showed of the voice agent, it seemed like it was generating the text and the audio at the same time.

90. 17:47

    No. Um, so the question was like, A, on the demo it looks like we are generating the text and the audio at the same time. Um,

91. 17:56

    no, for the, for the voice agent... Hello, Paul. It's me again. Can you say anything like, I don't know, a poem?

92. 18:06

    I'm afraid I can't recite poems.

93. 18:08

    It, it's just... So the text is produced in one go. It's just that I'm using a small LLM that is very fast, so it, it's nearly immediate and then the audio is, uh, is produced later.

94. 18:21

    Yeah. Yeah.

95. 18:23

    Uh, so thanks for a great model, by the way. Uh, I know that the weights are open. Is the voice cloning encoder also open?

96. 18:31

    Yeah, there is a small asterisk here. Um, we, we didn't release

97. 18:40

    this part, like the, the encoder part. Uh, which means it, it's the only thing that is missing for you to clone your own voice. Uh, it's a feature that we

98. 18:53

    only serve like, uh, in a proprietary fashion for now. So what you can do is use the text-to-speech model, use one of the open voices that we provide. Um, we may provide more, uh, in, in the future.

99. 19:09

    Uh, yeah, so far, we just didn't want to give everybody the ability to clone any voice.

100. 19:20

     Yeah.

101. 19:22

     Um, so, so big labs like, uh, Google and Mistral AI are working more on like a native voice-to-voice, let's say, models.

102. 19:30

     Yeah.

103. 19:31

     Your lab, ElevenLabs are doing more like this, like, cascading architecture. How ... Like what's your take on, uh, on both like guarding and how do you see things evolving?

104. 19:42

     What's my take? Um, so on the consumer side, you will always have the impression to speak to a single system that hear what you say and output something, right?

105. 19:56

     So it, it's purely an architectural, uh, model here. Um,

106. 20:02

     my take on this is that y- we can go very, very far by just using speech as an interface, uh, especially because these central LLM, they are extremely capable, but they also do a lot of things.

107. 20:17

     Uh, so just for the sake of being able to use any agent that has been released, uh, with the same interface, it has an advantage, uh, you know, to, uh, to interface.

108. 20:27

     So you, you can, you can go very far with just interface, especially if you are, uh, doing this kind of thing where you stream the, uh, the text token that's are output by the LLM.

109. 20:41

     Yeah. Last question because I'm out of time. Sorry.

110. 20:44

     Yeah. So the, the next steps on the interleaved, uh, function, which, uh, accepts both audio and text sounds really interesting. So if you do that real time, what do you see as the possibilities with that, uh, feature?

111. 20:58

     Um, no. So, so I, I, I didn't say that our next step would be this, right? I, I just said that there are several patterns to handle a stream of text as input as opposed to like a finite amount of text.

112. 21:13

     Uh, we actually don't know which one we'll, we'll choose. So whether it's interleaved or another solution like, uh, delayed s- uh, sequence modeling, for instance. Uh, so it's unclear which architecture is best, at least to, uh, to us, at least to me.

113. 21:28

     Uh, what it l- what it allows is lower latency because as soon as you have the first bit of text that are produced by the LLM, you can start voicing them out.

114. 21:39

     So you, you ... It-- In this agent, that was not clear because the, the utterance were very short. But imagine I ask Paul to generate a full page of text, um, it, it will be nice if I don't have to wait the end of the text generation to voice it out.

115. 21:58

     Yeah. Yeah. I mean, so, so it also increases kind of the UX side of things using these types of systems.

116. 22:03

     Yeah, absolutely.

117. 22:04

     Yeah.

118. 22:05

     Yeah.

119. 22:06

     Very nice. Thank you.

120. 22:07

     Thank you. [audience applauding] [outro music]